{
  "id": 678899,
  "title": "A Stitch in Time Saves Nine",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/678899",
  "author_name": "Adam Anderson",
  "post_date": "2026-02-25T17:28:50.696000",
  "votes": 36,
  "comment_count": 71,
  "views": 0,
  "content": "<p>If only the ancients who wrote these texts could imagine how difficult it would be to train computers to translate their work! Working with data from 1950-1750 BCE is not for the faint of heart. From 1927 CE until today, modern scholars have been working with the editing and publication of these ca. 8,000 documents in many different languages, and despite (or because of) almost 100 years of scholarly scrutiny, the datasets are more complicated and in a more convoluted format, than what the ancients originally wrote. These editions were migrated into a database, which became the foundation of the data we used for the training and supplemental data for this challenge. Unbeknownst to them, we are now re-purposing their editions for training ML models, and the messiness of the data is palpable for all to see. That said, this is the first benchmark for a significant ML application with the Old Assyrian data, so with great struggle comes greater pioneering outcomes.</p>\n<p>The purpose of this final update is to level the number of markers for gaps from two to one. The hope is that by making some small, but significant adjustments, this will improve the scores for everyone. What we refrained from doing was making any additional changes which could introduce new inconsistencies if not done uniformly. So for the final month of the competition, we will include a clear and concise list of recommendations for the training dataset for alignment with the test data.</p>\n<p><strong>Gap Change: Replaced all the following with a single <code>&lt;gap&gt;</code></strong></p>\n<ul>\n<li><code>x</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>[x]</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>…</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(break)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(large break)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(n broken lines)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>&lt;gap&gt;</code> <code>&lt;gap&gt;</code> —&gt; <code>&lt;gap&gt;</code></li>\n</ul>\n<p><strong>Results of these changes: no more <code>&lt;big_gap&gt;</code> and no more duplicates for <code>&lt;gap&gt;</code></strong> \nBefore the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single <code>&lt;gap&gt;</code>.</p>\n<table>\n<thead>\n<tr>\n<th>Gap</th>\n<th>Count</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt;</code></td>\n<td>30</td>\n</tr>\n<tr>\n<td><code>-&lt;gap&gt; &lt;gap&gt;</code></td>\n<td>9</td>\n</tr>\n<tr>\n<td><code>-&lt;gap&gt;-&lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt;-</code></td>\n<td>15</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;big_gap&gt;</code></td>\n<td>6</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt; &lt;gap&gt;</code></td>\n<td>2</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;big_gap&gt; &lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;gap&gt;-</code></td>\n<td>4</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;big_gap&gt;</code></td>\n<td>13</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt;</code></td>\n<td>15</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>12</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;gap&gt; &lt;big_gap&gt;</code></td>\n<td>2</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>3</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt; &lt;gap&gt; &lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;big_gap&gt; &lt;gap&gt; &lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>1</td>\n</tr>\n</tbody>\n</table>\n<p><strong>Alignment of Determinatives to match test data:</strong></p>\n<ul>\n<li><code>(d)</code> —&gt; <code>{d}</code></li>\n<li><code>(ki)</code> —&gt; <code>{ki}</code></li>\n<li><code>(TÚG)</code> —&gt; <code>TÚG</code></li>\n</ul>\n<p><strong>Shortening of long floats to four places after the decimal</strong></p>\n<ul>\n<li><code>1.3333300000000001</code> —&gt; <code>1.3333</code></li>\n<li><code>2.6666600000000003</code> —&gt; <code>2.6666</code></li>\n</ul>\n<p><strong>The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:</strong></p>\n<p>Some of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique <code>OARE_Text_ID</code> with the equivalent OARE IDs in the <code>published_texts.csv</code> dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (<a href=\"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)\" target=\"_blank\">https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)</a>.</p>\n<p><strong>Here are some recommended options for changes to the training data:</strong></p>\n<p><strong>Remove from translations:</strong></p>\n<ul>\n<li><code>fem.</code></li>\n<li><code>sing.</code></li>\n<li><code>pl.</code></li>\n<li><code>plural</code></li>\n<li><code>(?)</code></li>\n<li>any stray marks you find (e.g., <code>..</code>, <code>?</code>, <code>x</code>, <code>xx</code>, <code>&lt;&lt; &gt;&gt;</code>, <code>&lt; &gt;</code> except around <code>&lt;gap&gt;</code> of course)</li>\n<li>some of the translations equivocate two optional translations using <code>/</code>, it might be better to choose one or the other options provided, rather than including both with a <code>/</code> (e.g. \"you / she brought\" —&gt; \"you brought\" ).</li>\n</ul>\n<p><strong>Do not remove from translations (as these are in the test too):</strong></p>\n<ul>\n<li>quotation marks <code>\" \"</code></li>\n<li>apostrophes <code>'</code></li>\n<li>meaningful question marks <code>?</code> or exclamation marks <code>!</code></li>\n</ul>\n<p><strong>Replace in translations:</strong></p>\n<ul>\n<li><code>PN</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>-gold</code> —&gt; <code>pašallum gold</code></li>\n<li><code>-tax</code> —&gt; <code>šadduātum tax</code></li>\n<li><code>-textiles</code> —&gt; <code>kutānum textiles</code></li>\n<li><code>1 / 12 (shekel)</code> —&gt; <code>15 grains</code></li>\n<li><code>5 / 12 shekel</code> —&gt; <code>⅓ shekel 15 grains</code></li>\n<li><code>5 11 / 12 shekels</code> —&gt; <code>6 shekels less 15 grains</code></li>\n<li><code>7 / 12 shekel</code> —&gt; <code>½ shekel 15 grains</code></li>\n</ul>\n<p><strong>Decimals to Fractions:</strong></p>\n<ul>\n<li><code>0.5</code>  —&gt; <code>½</code></li>\n<li><code>0.25</code> —&gt; <code>¼</code></li>\n<li><code>0.3333</code> —&gt; <code>⅓</code></li>\n<li><code>0.8333</code> —&gt; <code>⅚</code></li>\n<li><code>0.625</code> —&gt; <code>⅝</code></li>\n<li><code>0.6666</code> —&gt; <code>⅔</code></li>\n<li><code>0.75</code> —&gt;  <code>¾</code></li>\n<li><code>0.1666</code> —&gt; <code>⅙</code></li>\n</ul>\n<p><strong>Change Roman numerals to integer numbers for months: e.g., month V —&gt; month 5</strong></p>\n<table>\n<thead>\n<tr>\n<th>Month</th>\n<th>Roman</th>\n<th>MN</th>\n<th>forms</th>\n<th>AKA</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Month 1</td>\n<td>I</td>\n<td>Bēlat ekallem</td>\n<td>be-el-tí-É.GAL-lim; be-el-té-kà-lim</td>\n<td>Bēlat-ekallim</td>\n</tr>\n<tr>\n<td>Month 2</td>\n<td>II</td>\n<td>ša-sarratim</td>\n<td>ša sá-ra-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 3</td>\n<td>III</td>\n<td>Kenātim</td>\n<td>ke-na-tim</td>\n<td>ša kēnātim</td>\n</tr>\n<tr>\n<td>Month 4</td>\n<td>IV</td>\n<td>Mahur-ilī</td>\n<td>Ma-hu-ur-DINGIR; ma-ḫu-ur-ì-lí</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 5</td>\n<td>V</td>\n<td>Abšarrani</td>\n<td>áb-ša-ra-ni; áb ša-ra-ni; áb-ša-ra-nu</td>\n<td>ab šarrāni; abšarrani</td>\n</tr>\n<tr>\n<td>Month 6</td>\n<td>VI</td>\n<td>Hubur</td>\n<td>Hu-bu-ur</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 7</td>\n<td>VII</td>\n<td>Ṣip'um</td>\n<td>ṣí-ip-im</td>\n<td>ṣipum</td>\n</tr>\n<tr>\n<td>Month 8</td>\n<td>VIII</td>\n<td>Qarrātum</td>\n<td>qá-ra-a-tí; qá-ra-a-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 9</td>\n<td>IX</td>\n<td>Kanwarta</td>\n<td>kán-bar-ta; Kà-an-ma-ar-ta</td>\n<td>Kanmarta</td>\n</tr>\n<tr>\n<td>Month 10</td>\n<td>X</td>\n<td>Te’inātum</td>\n<td>té-i-na-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 11</td>\n<td>XI</td>\n<td>Kuzallum</td>\n<td>ku-zal-li; ku-zal-lu</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 12</td>\n<td>XII</td>\n<td>Allanātum</td>\n<td>a-lá-na-tum; a-lá-na-tim</td>\n<td></td>\n</tr>\n</tbody>\n</table>\n<p><strong>Optional changes in transliterations:</strong></p>\n<ul>\n<li><code>Ḫ</code> → <code>H</code></li>\n<li><code>ḫ</code> → <code>h</code></li>\n<li><code>KÙ.B.</code> —&gt; <code>KÙ.BABBAR</code></li>\n</ul>\n<p><strong>Change unicode subscript numbers to normal integers in transliterations:</strong></p>\n<ul>\n<li><code>₀</code> → <code>0</code></li>\n<li><code>₁</code> → <code>1</code> </li>\n<li><code>₂</code> → <code>2</code></li>\n<li><code>₃</code> → <code>3</code></li>\n<li><code>₄</code> → <code>4</code></li>\n<li><code>₅</code> → <code>5</code></li>\n<li><code>₆</code> → <code>6</code></li>\n<li><code>₇</code> → <code>7</code></li>\n<li><code>₈</code> → <code>8</code></li>\n<li><code>₉</code> → <code>9</code></li>\n</ul>\n<p><strong>Decimals to Fractions for transliterations:</strong></p>\n<ul>\n<li><code>0.5</code>  —&gt; <code>½</code></li>\n<li><code>0.25</code> —&gt; <code>¼</code></li>\n<li><code>0.3333</code> —&gt; <code>⅓</code></li>\n<li><code>0.8333</code> —&gt; <code>⅚</code></li>\n<li><code>0.625</code> —&gt; <code>⅝</code></li>\n<li><code>0.6666</code> —&gt; <code>⅔</code></li>\n<li><code>0.75</code> —&gt;  <code>¾</code></li>\n<li><code>0.1666</code> —&gt; <code>⅙</code></li>\n</ul>\n<p><strong>Examples of the Outcomes for Test</strong></p>\n<table>\n<thead>\n<tr>\n<th>Transliteration</th>\n<th>Translation</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>1 e-ma-ar-šu &lt;gap&gt;</code></td>\n<td><code>1 donkey of his &lt;gap&gt;</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; a-na Ú-&lt;gap&gt; a-dí-in</code></td>\n<td><code>&lt;gap&gt; I gave it to U-&lt;gap&gt;.</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; ⅓ ma-na a-na En-na-nim DUMU Am-ri-a áš-qúl</code></td>\n<td><code>I paid &lt;gap&gt; ⅓ mina (silver) to Ennānum, son of Amriya.</code></td>\n</tr>\n</tbody>\n</table>\n<p><strong>Examples of the Optimal Outcomes for Train</strong></p>\n<table>\n<thead>\n<tr>\n<th>Transliteration</th>\n<th>Translation</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>&lt;gap&gt; ma-na KÙ.BABBAR ṣa-ru-pá-am &lt;gap&gt; GÍN KÙ.GI pá-ša-lam</code></td>\n<td><code>&lt;gap&gt; minas of refined silver, &lt;gap&gt; shekels of pašallum gold</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; GÚ SÍG.HI.A &lt;gap&gt; 5 maš-ku &lt;gap&gt; 22 na-ru-qá-tum 4 ANŠE ṣa-la-mu</code></td>\n<td><code>&lt;gap&gt; talents of wool, &lt;gap&gt; 5 hides, &lt;gap&gt; 22 sacks, 4 black donkeys</code></td>\n</tr>\n<tr>\n<td><code>⅓ ma-na 2 ½ GÍN KÙ.BABBAR 20 NINDA i-ṣé-er tù-wa-ra-a-ah-šu a-lá-hu-um i-šu</code></td>\n<td><code>Tuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.</code></td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": 3413989,
      "postDate": "2026-02-25T17:28:50.697Z",
      "content": "<p>If only the ancients who wrote these texts could imagine how difficult it would be to train computers to translate their work! Working with data from 1950-1750 BCE is not for the faint of heart. From 1927 CE until today, modern scholars have been working with the editing and publication of these ca. 8,000 documents in many different languages, and despite (or because of) almost 100 years of scholarly scrutiny, the datasets are more complicated and in a more convoluted format, than what the ancients originally wrote. These editions were migrated into a database, which became the foundation of the data we used for the training and supplemental data for this challenge. Unbeknownst to them, we are now re-purposing their editions for training ML models, and the messiness of the data is palpable for all to see. That said, this is the first benchmark for a significant ML application with the Old Assyrian data, so with great struggle comes greater pioneering outcomes.</p>\n<p>The purpose of this final update is to level the number of markers for gaps from two to one. The hope is that by making some small, but significant adjustments, this will improve the scores for everyone. What we refrained from doing was making any additional changes which could introduce new inconsistencies if not done uniformly. So for the final month of the competition, we will include a clear and concise list of recommendations for the training dataset for alignment with the test data.</p>\n<p><strong>Gap Change: Replaced all the following with a single <code>&lt;gap&gt;</code></strong></p>\n<ul>\n<li><code>x</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>[x]</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>…</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(break)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(large break)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>(n broken lines)</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>&lt;gap&gt;</code> <code>&lt;gap&gt;</code> —&gt; <code>&lt;gap&gt;</code></li>\n</ul>\n<p><strong>Results of these changes: no more <code>&lt;big_gap&gt;</code> and no more duplicates for <code>&lt;gap&gt;</code></strong> \nBefore the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single <code>&lt;gap&gt;</code>.</p>\n<table>\n<thead>\n<tr>\n<th>Gap</th>\n<th>Count</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt;</code></td>\n<td>30</td>\n</tr>\n<tr>\n<td><code>-&lt;gap&gt; &lt;gap&gt;</code></td>\n<td>9</td>\n</tr>\n<tr>\n<td><code>-&lt;gap&gt;-&lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt;-</code></td>\n<td>15</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;big_gap&gt;</code></td>\n<td>6</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;gap&gt; &lt;gap&gt;</code></td>\n<td>2</td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; &lt;big_gap&gt; &lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;gap&gt;-</code></td>\n<td>4</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;big_gap&gt;</code></td>\n<td>13</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt;</code></td>\n<td>15</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>12</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;gap&gt; &lt;big_gap&gt;</code></td>\n<td>2</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>3</td>\n</tr>\n<tr>\n<td><code>&lt;big_gap&gt; &lt;big_gap&gt; &lt;big_gap&gt; &lt;gap&gt; &lt;gap&gt;</code></td>\n<td>1</td>\n</tr>\n<tr>\n<td><code>-&lt;big_gap&gt; &lt;big_gap&gt; &lt;gap&gt; &lt;big_gap&gt; &lt;big_gap&gt;-</code></td>\n<td>1</td>\n</tr>\n</tbody>\n</table>\n<p><strong>Alignment of Determinatives to match test data:</strong></p>\n<ul>\n<li><code>(d)</code> —&gt; <code>{d}</code></li>\n<li><code>(ki)</code> —&gt; <code>{ki}</code></li>\n<li><code>(TÚG)</code> —&gt; <code>TÚG</code></li>\n</ul>\n<p><strong>Shortening of long floats to four places after the decimal</strong></p>\n<ul>\n<li><code>1.3333300000000001</code> —&gt; <code>1.3333</code></li>\n<li><code>2.6666600000000003</code> —&gt; <code>2.6666</code></li>\n</ul>\n<p><strong>The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:</strong></p>\n<p>Some of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique <code>OARE_Text_ID</code> with the equivalent OARE IDs in the <code>published_texts.csv</code> dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (<a href=\"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)\" target=\"_blank\">https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)</a>.</p>\n<p><strong>Here are some recommended options for changes to the training data:</strong></p>\n<p><strong>Remove from translations:</strong></p>\n<ul>\n<li><code>fem.</code></li>\n<li><code>sing.</code></li>\n<li><code>pl.</code></li>\n<li><code>plural</code></li>\n<li><code>(?)</code></li>\n<li>any stray marks you find (e.g., <code>..</code>, <code>?</code>, <code>x</code>, <code>xx</code>, <code>&lt;&lt; &gt;&gt;</code>, <code>&lt; &gt;</code> except around <code>&lt;gap&gt;</code> of course)</li>\n<li>some of the translations equivocate two optional translations using <code>/</code>, it might be better to choose one or the other options provided, rather than including both with a <code>/</code> (e.g. \"you / she brought\" —&gt; \"you brought\" ).</li>\n</ul>\n<p><strong>Do not remove from translations (as these are in the test too):</strong></p>\n<ul>\n<li>quotation marks <code>\" \"</code></li>\n<li>apostrophes <code>'</code></li>\n<li>meaningful question marks <code>?</code> or exclamation marks <code>!</code></li>\n</ul>\n<p><strong>Replace in translations:</strong></p>\n<ul>\n<li><code>PN</code> —&gt; <code>&lt;gap&gt;</code></li>\n<li><code>-gold</code> —&gt; <code>pašallum gold</code></li>\n<li><code>-tax</code> —&gt; <code>šadduātum tax</code></li>\n<li><code>-textiles</code> —&gt; <code>kutānum textiles</code></li>\n<li><code>1 / 12 (shekel)</code> —&gt; <code>15 grains</code></li>\n<li><code>5 / 12 shekel</code> —&gt; <code>⅓ shekel 15 grains</code></li>\n<li><code>5 11 / 12 shekels</code> —&gt; <code>6 shekels less 15 grains</code></li>\n<li><code>7 / 12 shekel</code> —&gt; <code>½ shekel 15 grains</code></li>\n</ul>\n<p><strong>Decimals to Fractions:</strong></p>\n<ul>\n<li><code>0.5</code>  —&gt; <code>½</code></li>\n<li><code>0.25</code> —&gt; <code>¼</code></li>\n<li><code>0.3333</code> —&gt; <code>⅓</code></li>\n<li><code>0.8333</code> —&gt; <code>⅚</code></li>\n<li><code>0.625</code> —&gt; <code>⅝</code></li>\n<li><code>0.6666</code> —&gt; <code>⅔</code></li>\n<li><code>0.75</code> —&gt;  <code>¾</code></li>\n<li><code>0.1666</code> —&gt; <code>⅙</code></li>\n</ul>\n<p><strong>Change Roman numerals to integer numbers for months: e.g., month V —&gt; month 5</strong></p>\n<table>\n<thead>\n<tr>\n<th>Month</th>\n<th>Roman</th>\n<th>MN</th>\n<th>forms</th>\n<th>AKA</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Month 1</td>\n<td>I</td>\n<td>Bēlat ekallem</td>\n<td>be-el-tí-É.GAL-lim; be-el-té-kà-lim</td>\n<td>Bēlat-ekallim</td>\n</tr>\n<tr>\n<td>Month 2</td>\n<td>II</td>\n<td>ša-sarratim</td>\n<td>ša sá-ra-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 3</td>\n<td>III</td>\n<td>Kenātim</td>\n<td>ke-na-tim</td>\n<td>ša kēnātim</td>\n</tr>\n<tr>\n<td>Month 4</td>\n<td>IV</td>\n<td>Mahur-ilī</td>\n<td>Ma-hu-ur-DINGIR; ma-ḫu-ur-ì-lí</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 5</td>\n<td>V</td>\n<td>Abšarrani</td>\n<td>áb-ša-ra-ni; áb ša-ra-ni; áb-ša-ra-nu</td>\n<td>ab šarrāni; abšarrani</td>\n</tr>\n<tr>\n<td>Month 6</td>\n<td>VI</td>\n<td>Hubur</td>\n<td>Hu-bu-ur</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 7</td>\n<td>VII</td>\n<td>Ṣip'um</td>\n<td>ṣí-ip-im</td>\n<td>ṣipum</td>\n</tr>\n<tr>\n<td>Month 8</td>\n<td>VIII</td>\n<td>Qarrātum</td>\n<td>qá-ra-a-tí; qá-ra-a-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 9</td>\n<td>IX</td>\n<td>Kanwarta</td>\n<td>kán-bar-ta; Kà-an-ma-ar-ta</td>\n<td>Kanmarta</td>\n</tr>\n<tr>\n<td>Month 10</td>\n<td>X</td>\n<td>Te’inātum</td>\n<td>té-i-na-tim</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 11</td>\n<td>XI</td>\n<td>Kuzallum</td>\n<td>ku-zal-li; ku-zal-lu</td>\n<td></td>\n</tr>\n<tr>\n<td>Month 12</td>\n<td>XII</td>\n<td>Allanātum</td>\n<td>a-lá-na-tum; a-lá-na-tim</td>\n<td></td>\n</tr>\n</tbody>\n</table>\n<p><strong>Optional changes in transliterations:</strong></p>\n<ul>\n<li><code>Ḫ</code> → <code>H</code></li>\n<li><code>ḫ</code> → <code>h</code></li>\n<li><code>KÙ.B.</code> —&gt; <code>KÙ.BABBAR</code></li>\n</ul>\n<p><strong>Change unicode subscript numbers to normal integers in transliterations:</strong></p>\n<ul>\n<li><code>₀</code> → <code>0</code></li>\n<li><code>₁</code> → <code>1</code> </li>\n<li><code>₂</code> → <code>2</code></li>\n<li><code>₃</code> → <code>3</code></li>\n<li><code>₄</code> → <code>4</code></li>\n<li><code>₅</code> → <code>5</code></li>\n<li><code>₆</code> → <code>6</code></li>\n<li><code>₇</code> → <code>7</code></li>\n<li><code>₈</code> → <code>8</code></li>\n<li><code>₉</code> → <code>9</code></li>\n</ul>\n<p><strong>Decimals to Fractions for transliterations:</strong></p>\n<ul>\n<li><code>0.5</code>  —&gt; <code>½</code></li>\n<li><code>0.25</code> —&gt; <code>¼</code></li>\n<li><code>0.3333</code> —&gt; <code>⅓</code></li>\n<li><code>0.8333</code> —&gt; <code>⅚</code></li>\n<li><code>0.625</code> —&gt; <code>⅝</code></li>\n<li><code>0.6666</code> —&gt; <code>⅔</code></li>\n<li><code>0.75</code> —&gt;  <code>¾</code></li>\n<li><code>0.1666</code> —&gt; <code>⅙</code></li>\n</ul>\n<p><strong>Examples of the Outcomes for Test</strong></p>\n<table>\n<thead>\n<tr>\n<th>Transliteration</th>\n<th>Translation</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>1 e-ma-ar-šu &lt;gap&gt;</code></td>\n<td><code>1 donkey of his &lt;gap&gt;</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; a-na Ú-&lt;gap&gt; a-dí-in</code></td>\n<td><code>&lt;gap&gt; I gave it to U-&lt;gap&gt;.</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; ⅓ ma-na a-na En-na-nim DUMU Am-ri-a áš-qúl</code></td>\n<td><code>I paid &lt;gap&gt; ⅓ mina (silver) to Ennānum, son of Amriya.</code></td>\n</tr>\n</tbody>\n</table>\n<p><strong>Examples of the Optimal Outcomes for Train</strong></p>\n<table>\n<thead>\n<tr>\n<th>Transliteration</th>\n<th>Translation</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><code>&lt;gap&gt; ma-na KÙ.BABBAR ṣa-ru-pá-am &lt;gap&gt; GÍN KÙ.GI pá-ša-lam</code></td>\n<td><code>&lt;gap&gt; minas of refined silver, &lt;gap&gt; shekels of pašallum gold</code></td>\n</tr>\n<tr>\n<td><code>&lt;gap&gt; GÚ SÍG.HI.A &lt;gap&gt; 5 maš-ku &lt;gap&gt; 22 na-ru-qá-tum 4 ANŠE ṣa-la-mu</code></td>\n<td><code>&lt;gap&gt; talents of wool, &lt;gap&gt; 5 hides, &lt;gap&gt; 22 sacks, 4 black donkeys</code></td>\n</tr>\n<tr>\n<td><code>⅓ ma-na 2 ½ GÍN KÙ.BABBAR 20 NINDA i-ṣé-er tù-wa-ra-a-ah-šu a-lá-hu-um i-šu</code></td>\n<td><code>Tuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.</code></td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "If only the ancients who wrote these texts could imagine how difficult it would be to train computers to translate their work! Working with data from 1950-1750 BCE is not for the faint of heart. From 1927 CE until today, modern scholars have been working with the editing and publication of these ca. 8,000 documents in many different languages, and despite (or because of) almost 100 years of scholarly scrutiny, the datasets are more complicated and in a more convoluted format, than what the ancients originally wrote. These editions were migrated into a database, which became the foundation of the data we used for the training and supplemental data for this challenge. Unbeknownst to them, we are now re-purposing their editions for training ML models, and the messiness of the data is palpable for all to see. That said, this is the first benchmark for a significant ML application with the Old Assyrian data, so with great struggle comes greater pioneering outcomes.\n\nThe purpose of this final update is to level the number of markers for gaps from two to one. The hope is that by making some small, but significant adjustments, this will improve the scores for everyone. What we refrained from doing was making any additional changes which could introduce new inconsistencies if not done uniformly. So for the final month of the competition, we will include a clear and concise list of recommendations for the training dataset for alignment with the test data.\n\n**Gap Change: Replaced all the following with a single `<gap>`**\n- `x` —> `<gap>`\n- `[x]` —> `<gap>`\n- `…` —> `<gap>`\n- `(break)` —> `<gap>`\n- `(large break)` —> `<gap>`\n- `(n broken lines)` —> `<gap>`\n- `<gap>` `<gap>` —> `<gap>`\n\n**Results of these changes: no more `<big_gap>` and no more duplicates for `<gap>`** \nBefore the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single `<gap>`.\n\n| Gap | Count |\n|-----|-------|\n|`<gap> <gap>`|30|\n|`-<gap> <gap>`|9|\n|`-<gap>-<gap>`|1|\n|`<gap> <gap>-`|15|\n|`<gap> <big_gap>`|6|\n|`<gap> <gap> <gap>`|2|\n|`<gap> <big_gap> <gap>`|1|\n|`-<big_gap> <gap>-`|4|\n|`-<big_gap> <big_gap>`|13|\n|`<big_gap> <big_gap>`|15|\n|`<big_gap> <big_gap>-`|12|\n|`<big_gap> <gap> <big_gap>`|2|\n|`<big_gap> <big_gap> <big_gap> <big_gap>-`|3|\n|`<big_gap> <big_gap> <big_gap> <gap> <gap>`|1|\n|`-<big_gap> <big_gap> <gap> <big_gap> <big_gap>-`|1|\n\n\n**Alignment of Determinatives to match test data:**\n- `(d)` —> `{d}`\n- `(ki)` —> `{ki}`\n- `(TÚG)` —> `TÚG`\n\n**Shortening of long floats to four places after the decimal**\n- `1.3333300000000001` —> `1.3333`\n- `2.6666600000000003` —> `2.6666`\n\n**The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:**\n\nSome of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique `OARE_Text_ID` with the equivalent OARE IDs in the `published_texts.csv` dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data).\n\n__Here are some recommended options for changes to the training data:__\n\n**Remove from translations:**\n- `fem.`\n- `sing.`\n- `pl.`\n- `plural`\n- `(?)`\n- any stray marks you find (e.g., `..`, `?`, `x`, `xx`, `<< >>`, `< >` except around `<gap>` of course)\n- some of the translations equivocate two optional translations using `/`, it might be better to choose one or the other options provided, rather than including both with a `/` (e.g. \"you / she brought\" —> \"you brought\" ).\n\n**Do not remove from translations (as these are in the test too):**\n- quotation marks `\" \"`\n- apostrophes `'`\n- meaningful question marks `?` or exclamation marks `!`\n\n**Replace in translations:**\n- `PN` —> `<gap>`\n- ` -gold` —> `pašallum gold`\n- ` -tax` —> `šadduātum tax`\n- ` -textiles` —> `kutānum textiles`\n- `1 / 12 (shekel)` —> `15 grains`\n- `5 / 12 shekel` —> `⅓ shekel 15 grains`\n- `5 11 / 12 shekels` —> `6 shekels less 15 grains`\n- `7 / 12 shekel` —> `½ shekel 15 grains`\n\n**Decimals to Fractions:**\n- `0.5`  —> `½`\n- `0.25` —> `¼`\n- `0.3333` —> `⅓`\n- `0.8333` —> `⅚`\n- `0.625` —> `⅝`\n- `0.6666` —> `⅔`\n- `0.75` —>  `¾`\n- `0.1666` —> `⅙`\n\n**Change Roman numerals to integer numbers for months: e.g., month V —> month 5**\n\n| Month | Roman | MN | forms | AKA |\n|--------|--------|-----|-------|------|\n|Month 1|\tI |\tBēlat ekallem | be-el-tí-É.GAL-lim; be-el-té-kà-lim |Bēlat-ekallim|\n|Month 2|\tII |\tša-sarratim |\tša sá-ra-tim|\t|\n|Month 3|\tIII |\tKenātim |\tke-na-tim |\tša kēnātim|\n|Month 4|\tIV |\tMahur-ilī |\tMa-hu-ur-DINGIR; ma-ḫu-ur-ì-lí |\t|\n|Month 5|\tV |\tAbšarrani |\táb-ša-ra-ni; áb ša-ra-ni; áb-ša-ra-nu |\tab šarrāni; abšarrani |\n|Month 6|\tVI |\tHubur |\tHu-bu-ur\t| |\n|Month 7|\tVII |\tṢip'um |\tṣí-ip-im |\tṣipum|\n|Month 8|\tVIII |\tQarrātum |\tqá-ra-a-tí; qá-ra-a-tim |\t|\n|Month 9|\tIX |\tKanwarta |\tkán-bar-ta; Kà-an-ma-ar-ta |\tKanmarta |\n|Month 10|\tX |\tTe’inātum |\tté-i-na-tim |\t|\n|Month 11|\tXI |\tKuzallum |\tku-zal-li; ku-zal-lu |\t|\n|Month 12|\tXII |\tAllanātum |\ta-lá-na-tum; a-lá-na-tim |\t|\n\n\n**Optional changes in transliterations:**\n- `Ḫ` → `H`\n- `ḫ` → `h`\n- `KÙ.B.` —> `KÙ.BABBAR`\n\n**Change unicode subscript numbers to normal integers in transliterations:**\n- `₀` → `0`\n- `₁` → `1` \n- `₂` → `2`\n- `₃` → `3`\n- `₄` → `4`\n- `₅` → `5`\n- `₆` → `6`\n- `₇` → `7`\n- `₈` → `8`\n- `₉` → `9`\n\n**Decimals to Fractions for transliterations:**\n- `0.5`  —> `½`\n- `0.25` —> `¼`\n- `0.3333` —> `⅓`\n- `0.8333` —> `⅚`\n- `0.625` —> `⅝`\n- `0.6666` —> `⅔`\n- `0.75` —>  `¾`\n- `0.1666` —> `⅙`\n\n**Examples of the Outcomes for Test**\n\n|Transliteration|Translation|\n|-----------|---------------|\n| `1 e-ma-ar-šu <gap>`| `1 donkey of his <gap>`|\n| `<gap> a-na Ú-<gap> a-dí-in`| `<gap> I gave it to U-<gap>.`|\n| `<gap> ⅓ ma-na a-na En-na-nim DUMU Am-ri-a áš-qúl`| `I paid <gap> ⅓ mina (silver) to Ennānum, son of Amriya.`|\n\n**Examples of the Optimal Outcomes for Train**\n\n|Transliteration|Translation|\n|-----------|---------------|\n|`<gap> ma-na KÙ.BABBAR ṣa-ru-pá-am <gap> GÍN KÙ.GI pá-ša-lam` | `<gap> minas of refined silver, <gap> shekels of pašallum gold` |\n|`<gap> GÚ SÍG.HI.A <gap> 5 maš-ku <gap> 22 na-ru-qá-tum 4 ANŠE ṣa-la-mu`|`<gap> talents of wool, <gap> 5 hides, <gap> 22 sacks, 4 black donkeys`|\n|`⅓ ma-na 2 ½ GÍN KÙ.BABBAR 20 NINDA i-ṣé-er tù-wa-ra-a-ah-šu a-lá-hu-um i-šu` | `Tuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.`|",
      "votes": 36
    },
    {
      "id": 3414072,
      "postDate": "2026-02-25T21:44:35.927Z",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a> Thanks for the update. Some clarification to be 101% sure:</p>\n<p><strong>About double quotes:</strong></p>\n<blockquote>\n  <p>Do not remove from translations (as these are in the test too):</p>\n  <p>quotation marks \" \"</p>\n</blockquote>\n<p>In the previous update:</p>\n<blockquote>\n  <p>quotations \" “= removed</p>\n</blockquote>\n<p>The final rule is keep all quotation marks, correct? And for curly double quotes, we have to convert them to plain double quotes or not?</p>\n<p><strong>About  parentheses:</strong></p>\n<p>In the previous update:</p>\n<blockquote>\n  <p>parentheses ( )= removed</p>\n</blockquote>\n<p>Now, we need to keep parenthesis as for the examples outcomes you've provided, right?</p>\n<p>For instance: <code>0faa50f7-b86c-466c-a8ab-3a6f48fcb00a</code></p>\n<p>\"at 52.5 grains (of silver) each, 0.5 shekel\" we must keep the parenthesis around <code>of silver</code>, right?</p>\n<p>We must not do: \"at 52.5 grains of silver each, 0.5 shekel\"</p>\n<p><strong>About replacements:</strong></p>\n<blockquote>\n  <p>-tax —&gt; šadduātum tax</p>\n  <p>-gold —&gt; pašallum gold</p>\n  <p>textiles —&gt; kutānum textiles (are you sure it's not -textiles —&gt; kutānum textiles )?</p>\n</blockquote>\n<p>And it's <code>textiles</code> and <code>-tax</code> and <code>-gold</code> surrounded by spaces? \nIf not then in <code>198e428d-f51b-40d1-96d8-aee4bfa60d8d</code>:</p>\n<p>\"of Ennam-Suen, 8 textiles as <strong>import-tax,</strong> 10 textiles are pre-emption\" </p>\n<p>would become: </p>\n<p>\"of Ennam-Suen, 8 kutānum textiles as <strong>importšadduātum tax</strong>, 10 kutānum textiles are pre-emption\". </p>\n<p>and it 229be03b-c772-4a6c-8792-c3f80948a97d:</p>\n<p>\"received 28 kutānu-textiles and 1 black donkey\"</p>\n<p>would become: </p>\n<p>\"received 28 <strong>kutānu-kutānum textiles</strong> and 1 black donkey\"</p>\n<p>Which would look weird, please confirm.</p>\n<p><strong>About <code>&lt;gap&gt;</code>:</strong></p>\n<p>In previous update:</p>\n<blockquote>\n  <p>added space around gap in translations (not transliterations)</p>\n</blockquote>\n<p>Is it still true? Do we need to add space around <code>&lt;gap&gt;</code> except if glued with a dash?</p>",
      "rawMarkdown": "@deeppast Thanks for the update. Some clarification to be 101% sure:\n\n**About double quotes:**\n\n>Do not remove from translations (as these are in the test too):\n\n>quotation marks \" \"\n\nIn the previous update:\n\n>quotations \" “= removed\n\nThe final rule is keep all quotation marks, correct? And for curly double quotes, we have to convert them to plain double quotes or not?\n\n**About  parentheses:**\n\nIn the previous update:\n>parentheses ( )= removed\n\nNow, we need to keep parenthesis as for the examples outcomes you've provided, right?\n\nFor instance: `0faa50f7-b86c-466c-a8ab-3a6f48fcb00a`\n\n\"at 52.5 grains (of silver) each, 0.5 shekel\" we must keep the parenthesis around `of silver`, right?\n\nWe must not do: \"at 52.5 grains of silver each, 0.5 shekel\"\n\n**About replacements:**\n>-tax —> šadduātum tax\n\n>-gold —> pašallum gold\n\n>textiles —> kutānum textiles (are you sure it's not -textiles —> kutānum textiles )?\n\nAnd it's `textiles` and `-tax` and `-gold` surrounded by spaces? \nIf not then in `198e428d-f51b-40d1-96d8-aee4bfa60d8d`:\n\n\"of Ennam-Suen, 8 textiles as **import-tax,** 10 textiles are pre-emption\" \n\nwould become: \n\n\"of Ennam-Suen, 8 kutānum textiles as **importšadduātum tax**, 10 kutānum textiles are pre-emption\". \n\nand it 229be03b-c772-4a6c-8792-c3f80948a97d:\n\n\"received 28 kutānu-textiles and 1 black donkey\"\n\nwould become: \n\n\"received 28 **kutānu-kutānum textiles** and 1 black donkey\"\n\nWhich would look weird, please confirm.\n\n**About `<gap>`:**\n\nIn previous update:\n>added space around gap in translations (not transliterations)\n\nIs it still true? Do we need to add space around `<gap>` except if glued with a dash?\n",
      "votes": 12,
      "replies": [
        {
          "id": 3414734,
          "postDate": "2026-02-27T14:50:56.837Z",
          "content": "<p>the questions about quotes and parentheses are extremely important</p>",
          "rawMarkdown": "the questions about quotes and parentheses are extremely important",
          "votes": 3
        },
        {
          "id": 3414893,
          "postDate": "2026-02-27T21:59:35.493Z",
          "content": "<p>Yes, keep parentheses and quotations should be dead quotes, no curl, same for apostrophes / scare quotes.</p>\n<blockquote>\n  <p>Do not remove from translations (as these are in the test too):\n  quotation marks <code>\" \"</code>\n  apostrophes <code>'</code>\n  meaningful question marks <code>?</code> or exclamation marks <code>!</code></p>\n</blockquote>\n<p>As for the words which begin with <code>-</code>: words which appear with an initial hyphen have a space before them, rather than those words which are joined to another word with a hyphen. These suggestions are meant to provide some context for the missing words, but the only way to know for sure is to check in the PDFs, which is why they were not replaced in v3. These missing parts were removed in the OARE database, and should be replaced in an ideal situation by checking the PDFs.</p>\n<ul>\n<li><code>-gold</code> —&gt; <code>pašallum gold</code></li>\n<li><code>-tax</code> —&gt; <code>šadduātum tax</code></li>\n<li><code>-textiles</code> —&gt; <code>kutānum textiles</code></li>\n</ul>",
          "rawMarkdown": "Yes, keep parentheses and quotations should be dead quotes, no curl, same for apostrophes / scare quotes.\n>Do not remove from translations (as these are in the test too):\n>quotation marks `\" \"`\n> apostrophes `'`\n>meaningful question marks `?` or exclamation marks `!`\n\nAs for the words which begin with `-`: words which appear with an initial hyphen have a space before them, rather than those words which are joined to another word with a hyphen. These suggestions are meant to provide some context for the missing words, but the only way to know for sure is to check in the PDFs, which is why they were not replaced in v3. These missing parts were removed in the OARE database, and should be replaced in an ideal situation by checking the PDFs.\n\n- ` -gold` —> `pašallum gold`\n- ` -tax` —> `šadduātum tax`\n- ` -textiles` —> `kutānum textiles`",
          "votes": 4
        }
      ]
    },
    {
      "id": 3416050,
      "postDate": "2026-03-02T00:10:52.533Z",
      "content": "<p>Thanks for the comprehensive update and all the clarifications in the comments. After applying the v3 changes, a few things that helped on my end:</p>\n<p>Recovering truncated transliterations: Confirmed that ~10% are affected. Matching oare_id against published_texts.csv and cross-referencing with the AKT PDFs recovered most of them. Be careful, though,  some entries aren't just truncated; the translation doesn't match the transliteration at all (as MPWARE pointed out for AKT 8, 55), so it's worth doing a length-ratio sanity check before blindly patching.</p>\n<p>The -textiles / -gold / -tax replacements: Based on the host's clarification, these only apply when preceded by a space (i.e., -textiles to kutānum textiles), not in the middle of a word, like import-tax or kutānu-textiles. A naive <em>str.replace</em> will break things, regex with word boundaries or explicit space matching is safer here.</p>\n<p>Straight quotes: Curly quotes to straight quotes for both \" and '. Small thing, but easy to miss if your text editor or PDF extraction reintroduces curly ones.</p>\n<p>Still working through the fraction conversions and subscript normalization. Good luck to everyone in the final stretch!</p>",
      "rawMarkdown": "Thanks for the comprehensive update and all the clarifications in the comments. After applying the v3 changes, a few things that helped on my end:\n\nRecovering truncated transliterations: Confirmed that ~10% are affected. Matching oare_id against published_texts.csv and cross-referencing with the AKT PDFs recovered most of them. Be careful, though,  some entries aren't just truncated; the translation doesn't match the transliteration at all (as MPWARE pointed out for AKT 8, 55), so it's worth doing a length-ratio sanity check before blindly patching.\n\nThe -textiles / -gold / -tax replacements: Based on the host's clarification, these only apply when preceded by a space (i.e., -textiles to kutānum textiles), not in the middle of a word, like import-tax or kutānu-textiles. A naive *str.replace* will break things, regex with word boundaries or explicit space matching is safer here.\n\nStraight quotes: Curly quotes to straight quotes for both \" and '. Small thing, but easy to miss if your text editor or PDF extraction reintroduces curly ones.\n\nStill working through the fraction conversions and subscript normalization. Good luck to everyone in the final stretch!",
      "votes": 7
    },
    {
      "id": 3414172,
      "postDate": "2026-02-26T06:35:02.263Z",
      "content": "<p>I think you should update these data from the beginning.Some samples from train.csv were aligned by hand. To revise these one more is really a challenging work.</p>",
      "rawMarkdown": "I think you should update these data from the beginning.Some samples from train.csv were aligned by hand. To revise these one more is really a challenging work.",
      "votes": 7,
      "replies": [
        {
          "id": 3414894,
          "postDate": "2026-02-27T21:59:51.867Z",
          "content": "<p>There will be no further updates, I'm sorry.</p>",
          "rawMarkdown": "There will be no further updates, I'm sorry.",
          "votes": 2
        }
      ]
    },
    {
      "id": 3414169,
      "postDate": "2026-02-26T06:28:22.013Z",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a>  hihi, </p>\n<p>Regarding the unit conversions that you mentioned below, I suspect that the right-hand-side of the first two rows should be interchanged</p>\n<pre><code>1 / 12 (shekel) —&gt; ⅔ shekel 15 grains\n5 / 12 shekel —&gt; 15 grains\n5 11 / 12 shekels —&gt; 6 shekels less 15 grains\n7 / 12 shekel —&gt; ½ shekel 15 grains\n</code></pre>\n<p>Is it that <code>1/12 shekel == 15 grains</code> ?</p>\n<p>if so, then the conversions should instead be </p>\n<pre><code>1 / 12 (shekel) —&gt; 15 grains\n5 / 12 shekel —&gt; ⅔ shekel 15 grains\n5 11 / 12 shekels —&gt; 6 shekels less 15 grains\n7 / 12 shekel —&gt; ½ shekel 15 grains\n</code></pre>\n<p>is it?</p>",
      "rawMarkdown": "@deeppast  hihi, \n\nRegarding the unit conversions that you mentioned below, I suspect that the right-hand-side of the first two rows should be interchanged\n\n```plaintext\n1 / 12 (shekel) —> ⅔ shekel 15 grains\n5 / 12 shekel —> 15 grains\n5 11 / 12 shekels —> 6 shekels less 15 grains\n7 / 12 shekel —> ½ shekel 15 grains\n```\n\nIs it that `1/12 shekel == 15 grains` ?\n\nif so, then the conversions should instead be \n\n```plaintext\n1 / 12 (shekel) —> 15 grains\n5 / 12 shekel —> ⅔ shekel 15 grains\n5 11 / 12 shekels —> 6 shekels less 15 grains\n7 / 12 shekel —> ½ shekel 15 grains\n```\n\nis it?",
      "votes": 6,
      "replies": [
        {
          "id": 3414891,
          "postDate": "2026-02-27T21:43:03.437Z",
          "content": "<p>close, 5 / 12 --&gt; ⅓ shekel 15 grains</p>",
          "rawMarkdown": "close, 5 / 12 --> ⅓ shekel 15 grains",
          "votes": 2
        },
        {
          "id": 3414997,
          "postDate": "2026-02-28T03:33:12.203Z",
          "content": "<p>Isn't it <code>5 / 12 shekel —&gt; ⅓ shekel 15 grains</code> instead of <code>5 / 12 shekel —&gt; ⅔ shekel 15 grains</code>?</p>\n<p>There are probably many similar mistakes in the test data.</p>",
          "rawMarkdown": "Isn't it `5 / 12 shekel —> ⅓ shekel 15 grains` instead of `5 / 12 shekel —> ⅔ shekel 15 grains`?\n\nThere are probably many similar mistakes in the test data.",
          "replies": [
            {
              "id": 3415162,
              "postDate": "2026-02-28T10:14:05.900Z",
              "content": "<p>oops nice catch you are right</p>",
              "rawMarkdown": "oops nice catch you are right"
            }
          ]
        }
      ]
    },
    {
      "id": 3414139,
      "postDate": "2026-02-26T03:27:06.307Z",
      "content": "<p>I am a bit confused about the table that follows 'Results of these changes: no more <code>&lt;big_gap&gt;</code>'. That table still says <code>&lt;gap&gt; &lt;gap&gt;: 30</code>, <code>&lt;gap&gt; &lt;big_gap&gt;: 6</code>, etc. Does this mean that there's till 30 occurrences of <code>&lt;gap&gt; &lt;gap&gt;</code>, 6 occurrences of <code>&lt;gap&gt; &lt;big_gap&gt;</code>, etc. in the test set?</p>\n<p>I thought I had understood that all of the <code>&lt;big_gap&gt;</code> had been replaced by <code>&lt;gap&gt;</code> and neighboring gaps had been merged together, but I'm less certain I understand right now.</p>\n<p>Edit: it seems that some time since this post <code>Results of these changes: no more &lt;big_gap&gt;</code> was changed to <code>Results of these changes: no more &lt;big_gap&gt; and no more duplicates for &lt;gap&gt; Before the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single &lt;gap&gt;.</code> so it's now clearer. Thank you for the clarification!</p>",
      "rawMarkdown": "I am a bit confused about the table that follows 'Results of these changes: no more `<big_gap>`'. That table still says `<gap> <gap>: 30`, `<gap> <big_gap>: 6`, etc. Does this mean that there's till 30 occurrences of `<gap> <gap>`, 6 occurrences of `<gap> <big_gap>`, etc. in the test set?\n\nI thought I had understood that all of the `<big_gap>` had been replaced by `<gap>` and neighboring gaps had been merged together, but I'm less certain I understand right now.\n\nEdit: it seems that some time since this post `Results of these changes: no more <big_gap>` was changed to `Results of these changes: no more <big_gap> and no more duplicates for <gap> Before the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single <gap>.` so it's now clearer. Thank you for the clarification!",
      "votes": 6,
      "replies": [
        {
          "id": 3414146,
          "postDate": "2026-02-26T03:47:32.113Z",
          "content": "<p>In this very post, i can't find any info about merging gap…</p>",
          "rawMarkdown": "In this very post, i can't find any info about merging gap...",
          "replies": [
            {
              "id": 3414148,
              "postDate": "2026-02-26T03:53:39.457Z",
              "content": "<p>That's what had been mentioned in last week's update:</p>\n<pre><code>Gaps, damage markers, and parallel alignment [update: 2/18/26]\n\nAnother recurring source of confusion concerns damaged text and gap markers in the data:\n\n    x represents a single broken sign,\n    sequences like x x x x or ... represent a larger lacuna.\n\nFor modeling purposes we reduced all breaks to a single marker: &lt;gap&gt;\n\n we removed the tag for &lt;big_gap&gt; from the train and test (and other transliterations). We also deduplicated instances multiple sequential gaps (e.g. &lt;gap&gt; &lt;gap, &lt;gap&gt;-&lt;gap&gt;, &lt;gap&gt; &lt;gap&gt;, &lt;gap&gt;. &lt;gap&gt;, etc.\n</code></pre>\n<p>Source: <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209\" target=\"_blank\">https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209</a></p>",
              "rawMarkdown": "That's what had been mentioned in last week's update:\n\n```\nGaps, damage markers, and parallel alignment [update: 2/18/26]\n\nAnother recurring source of confusion concerns damaged text and gap markers in the data:\n\n    x represents a single broken sign,\n    sequences like x x x x or ... represent a larger lacuna.\n\nFor modeling purposes we reduced all breaks to a single marker: <gap>\n\n we removed the tag for <big_gap> from the train and test (and other transliterations). We also deduplicated instances multiple sequential gaps (e.g. <gap> <gap, <gap>-<gap>, <gap> <gap>, <gap>. <gap>, etc.\n```\nSource: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209"
            },
            {
              "id": 3414149,
              "postDate": "2026-02-26T03:58:27.463Z",
              "content": "<p>thanks, i'm aware, just thought this post is a comprehensive summary of all the changes, which apparently is not…</p>",
              "rawMarkdown": "thanks, i'm aware, just thought this post is a comprehensive summary of all the changes, which apparently is not...",
              "votes": 2
            }
          ]
        }
      ]
    },
    {
      "id": 3418958,
      "postDate": "2026-03-09T15:51:14.303Z",
      "content": "<p>Sorry if I am being dumb here</p>\n<pre><code>The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:\n</code></pre>\n<p>why is it “no guarantee”? IMO these are deterministic postprocessing, so if the test set conforms to this, the score will 100% improve. Does this mean only some texts were changed in this manner while the others are left as it is?</p>",
      "rawMarkdown": "Sorry if I am being dumb here\n\n```\nThe rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:\n```\n\nwhy is it “no guarantee”? IMO these are deterministic postprocessing, so if the test set conforms to this, the score will 100% improve. Does this mean only some texts were changed in this manner while the others are left as it is?",
      "votes": 3,
      "replies": [
        {
          "id": 3419052,
          "postDate": "2026-03-09T19:29:08.723Z",
          "content": "<p>Others have pointed that some of those recommendations decrease LB score. I'm guessing either the test set is inconsistent, or it doens't respect those guidelines at all.</p>",
          "rawMarkdown": "Others have pointed that some of those recommendations decrease LB score. I'm guessing either the test set is inconsistent, or it doens't respect those guidelines at all.",
          "votes": 3
        }
      ]
    },
    {
      "id": 3415113,
      "postDate": "2026-02-28T07:33:26.507Z",
      "content": "<p>It looks like the “Change Roman numerals to integer numbers for months” table is misaligned/broken: Month 2 is labeled as Roman III, and III appears twice. </p>\n<p>According to Kouwenberg (Introduction to Old Assyrian, p. 182), the Old Assyrian calendar starts with Month 1 = I (Bēlat ekallem/Bēltekallem), followed by Month 2 = II (ša sarrātim) and Month 3 = III (ša kēnātim).\nCould you add Month 1 and shift/correct the Roman numerals accordingly?</p>",
      "rawMarkdown": "It looks like the “Change Roman numerals to integer numbers for months” table is misaligned/broken: Month 2 is labeled as Roman III, and III appears twice. \n\nAccording to Kouwenberg (Introduction to Old Assyrian, p. 182), the Old Assyrian calendar starts with Month 1 = I (Bēlat ekallem/Bēltekallem), followed by Month 2 = II (ša sarrātim) and Month 3 = III (ša kēnātim).\nCould you add Month 1 and shift/correct the Roman numerals accordingly?",
      "votes": 3
    },
    {
      "id": 3414479,
      "postDate": "2026-02-26T21:40:52.667Z",
      "content": "<p>The new update seems to have truncated the longer transliteration strings. The previous training data had the following long training entries:</p>\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>592</td>\n      <td>746</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>338</td>\n      <td>588</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>343</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>408</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>357</td>\n      <td>626</td>\n    </tr>\n    <tr>\n      <td>1260</td>\n      <td>c97bb594-a5a1-4674-9496-48496e91c2ee</td>\n      <td>29</td>\n      <td>602</td>\n    </tr>\n    <tr>\n      <td>1378</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>379</td>\n      <td>603</td>\n    </tr>\n  </tbody>\n</table>\n<p><br>\n<strong>The new data has truncated these to be:</strong></p>\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>138</td>\n      <td>739</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>107</td>\n      <td>587</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>114</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>108</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>107</td>\n      <td>625</td>\n    </tr>\n    <tr>\n      <td>1377</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>120</td>\n      <td>585</td>\n    </tr>\n  </tbody>\n</table>\n<p><br>\nI'm not sure if this applies to the other rows too, but it certainly seems to have lost a lot of data.</p>",
      "rawMarkdown": "The new update seems to have truncated the longer transliteration strings. The previous training data had the following long training entries:\n\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>592</td>\n      <td>746</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>338</td>\n      <td>588</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>343</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>408</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>357</td>\n      <td>626</td>\n    </tr>\n    <tr>\n      <td>1260</td>\n      <td>c97bb594-a5a1-4674-9496-48496e91c2ee</td>\n      <td>29</td>\n      <td>602</td>\n    </tr>\n    <tr>\n      <td>1378</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>379</td>\n      <td>603</td>\n    </tr>\n  </tbody>\n</table>\n\n<br>\n**The new data has truncated these to be:**\n\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>138</td>\n      <td>739</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>107</td>\n      <td>587</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>114</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>108</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>107</td>\n      <td>625</td>\n    </tr>\n    <tr>\n      <td>1377</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>120</td>\n      <td>585</td>\n    </tr>\n  </tbody>\n</table>\n\n\n<br>\nI'm not sure if this applies to the other rows too, but it certainly seems to have lost a lot of data.",
      "votes": 3,
      "replies": [
        {
          "id": 3414484,
          "postDate": "2026-02-26T22:03:52.923Z",
          "content": "<p>Around 10% of transliterations are truncated. We've to fix that ourself according to the instructions.</p>",
          "rawMarkdown": "Around 10% of transliterations are truncated. We've to fix that ourself according to the instructions.",
          "votes": 1,
          "replies": [
            {
              "id": 3414560,
              "postDate": "2026-02-27T05:52:55.683Z",
              "content": "<p>which instructions?</p>",
              "rawMarkdown": "which instructions?",
              "votes": 1
            },
            {
              "id": 3414580,
              "postDate": "2026-02-27T07:36:24.560Z",
              "content": "<blockquote>\n  <p>Some of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique OARE_Text_ID with the equivalent OARE IDs in the published_texts.csv dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (<a href=\"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)\" target=\"_blank\">https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)</a>.</p>\n</blockquote>",
              "rawMarkdown": ">Some of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique OARE_Text_ID with the equivalent OARE IDs in the published_texts.csv dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data).",
              "votes": 1
            },
            {
              "id": 3414593,
              "postDate": "2026-02-27T08:18:10.987Z",
              "content": "<p>It's a bit bad though that data that was there, and which I spent a large amount of time working out how to split, has now suddenly disappeared from the data set. </p>",
              "rawMarkdown": "It's a bit bad though that data that was there, and which I spent a large amount of time working out how to split, has now suddenly disappeared from the data set. ",
              "votes": 5
            }
          ]
        },
        {
          "id": 3415110,
          "postDate": "2026-02-28T07:26:44.910Z",
          "content": "<p>A large number of data where the ratio between the lengths of transliteration and translation seem weird have wrong/truncated transliteration/translation, e.g. 8376cbda-b423-42d4-abb5-188d04896392. Can follow the instruction to find the raw pdf and amend these data case by case.</p>",
          "rawMarkdown": "A large number of data where the ratio between the lengths of transliteration and translation seem weird have wrong/truncated transliteration/translation, e.g. 8376cbda-b423-42d4-abb5-188d04896392. Can follow the instruction to find the raw pdf and amend these data case by case."
        }
      ]
    },
    {
      "id": 3416596,
      "postDate": "2026-03-03T10:48:47.297Z",
      "content": "<p>question about (n broken lines), do the lines between \"Nimar-Istar.\" and \"When he returns from Mamma he will bring\nit to me\" counts as n broken lines?</p>\n<p>If I understand correctly, the translation should be:</p>\n<p>I furthermore gave 1 mina of good, native copper for an allu’āru-container of sweet wine from Mamma to Puzur-Amurrum son of Nimar-Ištar. &lt;gap&gt; When he returns from Mamma he will bring it to me. (This was) apart from the 4 allu’āru-containers, the proceeds from the silver and washed copper that they have received (previously). Witnessed by Šu-Bēlum son of Kuzizia.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21691005%2F5a2fcafea533880783e26906cb8cdffd%2FScreenshot%202026-03-03%20174635.png?generation=1772534814707795&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "question about (n broken lines), do the lines between \"Nimar-Istar.\" and \"When he returns from Mamma he will bring\nit to me\" counts as n broken lines?\n\nIf I understand correctly, the translation should be:\n\nI furthermore gave 1 mina of good, native copper for an allu’āru-container of sweet wine from Mamma to Puzur-Amurrum son of Nimar-Ištar. \\<gap> When he returns from Mamma he will bring it to me. (This was) apart from the 4 allu’āru-containers, the proceeds from the silver and washed copper that they have received (previously). Witnessed by Šu-Bēlum son of Kuzizia.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21691005%2F5a2fcafea533880783e26906cb8cdffd%2FScreenshot%202026-03-03%20174635.png?generation=1772534814707795&alt=media)",
      "votes": 1,
      "replies": [
        {
          "id": 3419308,
          "postDate": "2026-03-10T12:22:04.547Z",
          "content": "<p>I think in this case there is no gap in the translation, it just appears that way because of the alignment.</p>",
          "rawMarkdown": "I think in this case there is no gap in the translation, it just appears that way because of the alignment.",
          "votes": 2
        },
        {
          "id": 3421590,
          "postDate": "2026-03-15T22:05:18.830Z",
          "content": "<p>Where are you getting these translations from?</p>",
          "rawMarkdown": "Where are you getting these translations from?",
          "replies": [
            {
              "id": 3421594,
              "postDate": "2026-03-15T22:24:23.353Z",
              "content": "<p>It's the PDFs from the second link here: <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/668485\" target=\"_blank\">https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/668485</a></p>",
              "rawMarkdown": "It's the PDFs from the second link here: https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/668485"
            }
          ]
        }
      ]
    },
    {
      "id": 3414925,
      "postDate": "2026-02-27T23:53:45.343Z",
      "content": "<p>Regarding the terms \"pašallum gold\", \"šadduātum tax\", \"kutānum textiles\", those appear in different forms in the texts. Are those the forms that are expected from the test data (e.g. \"kutanum-textile\", \"kutanu textile\" -&gt; \"kutānum textiles\")?</p>",
      "rawMarkdown": "Regarding the terms \"pašallum gold\", \"šadduātum tax\", \"kutānum textiles\", those appear in different forms in the texts. Are those the forms that are expected from the test data (e.g. \"kutanum-textile\", \"kutanu textile\" -> \"kutānum textiles\")?",
      "votes": 1,
      "replies": [
        {
          "id": 3414939,
          "postDate": "2026-02-28T00:55:34.600Z",
          "content": "<p>I don’t think that’s the case. </p>\n<p>First, words like pašallum, šadduātum, and kutānum do not exist in train.csv at all. </p>\n<p>Second, submissions that do not include words such as pašallum, šadduātum, and kutānum have achieved higher scores.</p>\n<p>The host only seemed to suggest that we do this. :)</p>",
          "rawMarkdown": "I don’t think that’s the case. \n\nFirst, words like pašallum, šadduātum, and kutānum do not exist in train.csv at all. \n\nSecond, submissions that do not include words such as pašallum, šadduātum, and kutānum have achieved higher scores.\n\nThe host only seemed to suggest that we do this. :)\n ",
          "votes": 4,
          "replies": [
            {
              "id": 3414946,
              "postDate": "2026-02-28T01:36:26.513Z",
              "content": "<p>when i drop the '[', ['] in the train, my scores would drop.  Normally, if dropping the \"[\" and \"]\" , the scores would have been improved. I don't know why it is like this.</p>",
              "rawMarkdown": "when i drop the '[', ['] in the train, my scores would drop.  Normally, if dropping the \"[\" and \"]\" , the scores would have been improved. I don't know why it is like this.",
              "votes": 1
            },
            {
              "id": 3415026,
              "postDate": "2026-02-28T04:35:10.683Z",
              "content": "<p>In your transliterations or translations? </p>",
              "rawMarkdown": "In your transliterations or translations? "
            }
          ]
        }
      ]
    },
    {
      "id": 3414179,
      "postDate": "2026-02-26T06:44:41.323Z",
      "content": "<p>just for the record, <strong>this update does not include all the changes made in the previous update</strong><br>\nFor example unicode subscript numbers are now present in the train transliteration:<br>\n<code>0064939c-59b9-4448-a63d-34612af0a1b5 -&gt; 1 TÚG ša qá-tim i-tur₄-DINGIR il₅-qé ...</code></p>",
      "rawMarkdown": "just for the record, **this update does not include all the changes made in the previous update**  \nFor example unicode subscript numbers are now present in the train transliteration:   \t\n`0064939c-59b9-4448-a63d-34612af0a1b5 -> 1 TÚG ša qá-tim i-tur₄-DINGIR il₅-qé ...`",
      "votes": 1
    },
    {
      "id": 3413998,
      "postDate": "2026-02-25T17:50:26.173Z",
      "content": "<p>Thanks, these are some new changes. Do we expect another test set update?</p>",
      "rawMarkdown": "Thanks, these are some new changes. Do we expect another test set update?",
      "votes": 1,
      "replies": [
        {
          "id": 3414001,
          "postDate": "2026-02-25T18:01:19.850Z",
          "content": "<p>Yes, both test and train were updated (final update 2/26/26).</p>",
          "rawMarkdown": "Yes, both test and train were updated (final update 2/26/26)."
        }
      ]
    },
    {
      "id": 3415275,
      "postDate": "2026-02-28T15:21:30.847Z",
      "content": "<p>I'm trying to identify the mapping between oare_id in train.csv and the AKT PDF in order to complete broken sections, I've found 1561 matches:</p>\n<ul>\n<li>AKT 5 = 77 oare_id </li>\n<li>AKT 6a = 304 oare_id</li>\n<li>AKT 6b = 222 oare_id</li>\n<li>AKT 6c = 201 oare_id</li>\n<li>AKT 6d = 139 oare_id</li>\n<li>AKT 6e = 255 oare_id</li>\n<li>AKT 8 = 363 oare_id (e.g. <code>0123a9b9-e81e-4d7a-a79b-10e7c0aacbb9</code> Kt 91/k 471, page 213)</li>\n</ul>\n<p>Someone with similar results? <a href=\"https://www.kaggle.com/jackvd\" target=\"_blank\">@jackvd</a> You was saying around 400 items was broken?</p>",
      "rawMarkdown": "I'm trying to identify the mapping between oare_id in train.csv and the AKT PDF in order to complete broken sections, I've found 1561 matches:\n- AKT 5 = 77 oare_id \n- AKT 6a = 304 oare_id\n- AKT 6b = 222 oare_id\n- AKT 6c = 201 oare_id\n- AKT 6d = 139 oare_id\n- AKT 6e = 255 oare_id\n- AKT 8 = 363 oare_id (e.g. `0123a9b9-e81e-4d7a-a79b-10e7c0aacbb9` Kt 91/k 471, page 213)\n\nSomeone with similar results? @jackvd You was saying around 400 items was broken?",
      "votes": 2,
      "replies": [
        {
          "id": 3415360,
          "postDate": "2026-02-28T18:50:07.417Z",
          "content": "<p>I have similar results with regard to AKT 5 and AKT 6a. I still have to verify a few ids to be sure but when I do I can post it in the updated <code>train.csv</code> dataset to make everyone lives a bit easier.</p>\n<p>Edit: The <a href=\"https://www.kaggle.com/datasets/angantyr/updated-train-w-publication-source-96-7-cov/data\" target=\"_blank\">dataset with sources included</a> should be up and running.</p>\n<table>\n<thead>\n<tr>\n<th>Publication</th>\n<th>page</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>AKT 5</td>\n<td>76</td>\n</tr>\n<tr>\n<td>AKT 6a</td>\n<td>295</td>\n</tr>\n<tr>\n<td>AKT 6b</td>\n<td>218</td>\n</tr>\n<tr>\n<td>AKT 6c</td>\n<td>196</td>\n</tr>\n<tr>\n<td>AKT 6d</td>\n<td>137</td>\n</tr>\n<tr>\n<td>AKT 6e</td>\n<td>233</td>\n</tr>\n<tr>\n<td>AKT 8</td>\n<td>355</td>\n</tr>\n</tbody>\n</table>",
          "rawMarkdown": "I have similar results with regard to AKT 5 and AKT 6a. I still have to verify a few ids to be sure but when I do I can post it in the updated `train.csv` dataset to make everyone lives a bit easier.\n\nEdit: The [dataset with sources included](https://www.kaggle.com/datasets/angantyr/updated-train-w-publication-source-96-7-cov/data) should be up and running.\n\n| Publication | page |\n| --- | --- |\n| AKT 5 | 76 |\n| AKT 6a | 295 |\n| AKT 6b | 218 |\n| AKT 6c | 196 |\n| AKT 6d | 137 |\n| AKT 6e | 233 |\n| AKT 8 | 355 |",
          "votes": 6,
          "replies": [
            {
              "id": 3415774,
              "postDate": "2026-03-01T09:50:59.623Z",
              "content": "<p>I've updated my list:</p>\n<pre><code>pub_pd = pd.read_csv(\"data/published_texts.csv\")\n\ntrain_df = pd.read_csv(\"data/train.csv\")\n\nakt_pd = pub_pd[(pub_pd['oare_id'].isin(train_df[\"oare_id\"].unique()))][[\"oare_id\", \"label\", \"excavation_no\", \"transliteration\"]]\nakt_pd[\"pdf\"] = akt_pd[\"label\"].str.extract(\n    r\"\\bAKT\\s+(\\d+[a-zA-Z]?)\",\n    expand=False\n)\n</code></pre>\n<pre><code>pdf\n5      77\n6a    304\n6b    222\n6c    201\n6d    139\n6e    255\n8     363\n</code></pre>\n<p>Some rows in train.csv are more than broken, the translation just does not match the transliteration at all. For instance: AKT8, 55. Kt 91/k 304 (1-161-91) oare_id: <code>5f088d12-ed99-434a-a113-65deab7e1426</code></p>\n<p><strong>In the PDF:</strong>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F11e1483f1b5212cfa6310549bc4f6a4e%2Fk304.png?generation=1772371430299359&amp;alt=media\" alt=\"\"></p>\n<p><strong>In train.csv:</strong> whatever v1,v2 or v3\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F838802a47acb10736ea3c509af8d3116%2Fk304-bad.png?generation=1772371460757455&amp;alt=media\" alt=\"\"></p>\n<p><strong>After RE-OCR:</strong>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2Fb94e2b6c02e584b71d68d4246257ccb9%2Fk304-fixed.png?generation=1772372008286642&amp;alt=media\" alt=\"\"></p>\n<p>oare_ids from AKT 6e are one of the most broken.</p>\n<p>This competition is about normalization and cleaning.</p>",
              "rawMarkdown": "I've updated my list:\n\n```python\npub_pd = pd.read_csv(\"data/published_texts.csv\")\n\ntrain_df = pd.read_csv(\"data/train.csv\")\n\nakt_pd = pub_pd[(pub_pd['oare_id'].isin(train_df[\"oare_id\"].unique()))][[\"oare_id\", \"label\", \"excavation_no\", \"transliteration\"]]\nakt_pd[\"pdf\"] = akt_pd[\"label\"].str.extract(\n    r\"\\bAKT\\s+(\\d+[a-zA-Z]?)\",\n    expand=False\n)\n```\n```python\npdf\n5      77\n6a    304\n6b    222\n6c    201\n6d    139\n6e    255\n8     363\n```\n\nSome rows in train.csv are more than broken, the translation just does not match the transliteration at all. For instance: AKT8, 55. Kt 91/k 304 (1-161-91) oare_id: `5f088d12-ed99-434a-a113-65deab7e1426`\n\n**In the PDF:**\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F11e1483f1b5212cfa6310549bc4f6a4e%2Fk304.png?generation=1772371430299359&alt=media)\n\n**In train.csv:** whatever v1,v2 or v3\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F838802a47acb10736ea3c509af8d3116%2Fk304-bad.png?generation=1772371460757455&alt=media)\n\n**After RE-OCR:**\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2Fb94e2b6c02e584b71d68d4246257ccb9%2Fk304-fixed.png?generation=1772372008286642&alt=media)\n\noare_ids from AKT 6e are one of the most broken.\n\nThis competition is about normalization and cleaning.",
              "votes": 5
            },
            {
              "id": 3416194,
              "postDate": "2026-03-02T09:25:40.107Z",
              "content": "<p>I wonder for how many of such cases we could employ a set match filtering, e.g., 'a-na'/'from', 'ku.babbar'/'silver', etc.</p>\n<p>It would not be a guarantee but a first stage flag to have a manual check and there are many words that have a 1:1 equivalent with English.</p>",
              "rawMarkdown": "I wonder for how many of such cases we could employ a set match filtering, e.g., 'a-na'/'from', 'ku.babbar'/'silver', etc.\n\nIt would not be a guarantee but a first stage flag to have a manual check and there are many words that have a 1:1 equivalent with English."
            }
          ]
        }
      ]
    },
    {
      "id": 3414364,
      "postDate": "2026-02-26T16:02:45.813Z",
      "content": "<p>Can't describe an amount of confusion when comparing a new train dropped last week and a new train from yesterday and now I'm trying do understand how to merge those two</p>\n<p>upd: sorry, my humble verdict is that these updates last week made things worse…</p>",
      "rawMarkdown": "Can't describe an amount of confusion when comparing a new train dropped last week and a new train from yesterday and now I'm trying do understand how to merge those two\n\nupd: sorry, my humble verdict is that these updates last week made things worse...",
      "votes": 2,
      "replies": [
        {
          "id": 3414370,
          "postDate": "2026-02-26T16:14:06.147Z",
          "content": "<p>how to understand this:</p>\n<pre><code>before: 31 kutānu-textiles of Iddin-Suen, 30 kutānu-textiles of Ah-šalim, 46 kutānu-textiles of hinnāya and Uṣurānum, 21 kutānu-textiles of Aššur-rēī, 36 kutānu-textiles of Šu-Ištar, 28 kutānu-textiles of Ennam-Suen, 24 kutānu-textiles of Lā-qēp, 18 kutānu-textiles of the merchant, 7 kutānu-textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\nafter:  31 -textiles of Iddin-Suen, 30 -textiles of Ah-šalim, 46 -textiles of hinnāya and Uṣurānum, 21 -textiles of Aššur-rē'ī, 36 -textiles of Šu-Ištar, 28 -textiles of Ennam-Suen, 24 -textiles of Lā-qēp, 18 -textiles of the merchant, 7 -textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\n</code></pre>",
          "rawMarkdown": "how to understand this:\n\n```\nbefore: 31 kutānu-textiles of Iddin-Suen, 30 kutānu-textiles of Ah-šalim, 46 kutānu-textiles of hinnāya and Uṣurānum, 21 kutānu-textiles of Aššur-rēī, 36 kutānu-textiles of Šu-Ištar, 28 kutānu-textiles of Ennam-Suen, 24 kutānu-textiles of Lā-qēp, 18 kutānu-textiles of the merchant, 7 kutānu-textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\nafter:  31 -textiles of Iddin-Suen, 30 -textiles of Ah-šalim, 46 -textiles of hinnāya and Uṣurānum, 21 -textiles of Aššur-rē'ī, 36 -textiles of Šu-Ištar, 28 -textiles of Ennam-Suen, 24 -textiles of Lā-qēp, 18 -textiles of the merchant, 7 -textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\n```",
          "replies": [
            {
              "id": 3414465,
              "postDate": "2026-02-26T20:43:13.670Z",
              "content": "<p>Yes, you can see in the post above, I recommend making some small changes, egg. <code>-textiles</code> --&gt; <code>kutānum textiles</code>, but this should be checked with the AKT volumes in PDF to be certain. The interum data (v2) update did this, but it introduced new errors because of a simple search / replace. Those changes were rolled back in the lates update (v3), and instructions were provided how to deal with those (above).</p>",
              "rawMarkdown": "Yes, you can see in the post above, I recommend making some small changes, egg. `-textiles` --> `kutānum textiles`, but this should be checked with the AKT volumes in PDF to be certain. The interum data (v2) update did this, but it introduced new errors because of a simple search / replace. Those changes were rolled back in the lates update (v3), and instructions were provided how to deal with those (above).",
              "votes": -1
            },
            {
              "id": 3414732,
              "postDate": "2026-02-27T14:49:29.263Z",
              "content": "<p>do you recommend to take V3 and manually update it based on the tutorial or to take V2 and just fix 's and quotes?</p>",
              "rawMarkdown": "do you recommend to take V3 and manually update it based on the tutorial or to take V2 and just fix 's and quotes?"
            },
            {
              "id": 3414898,
              "postDate": "2026-02-27T22:22:59.053Z",
              "content": "<p>Maybe even v1 is the least flawed, especially if you have already done the work to fix the gaps on it.</p>",
              "rawMarkdown": "Maybe even v1 is the least flawed, especially if you have already done the work to fix the gaps on it.",
              "votes": 2
            },
            {
              "id": 3414927,
              "postDate": "2026-02-27T23:58:08.460Z",
              "content": "<p>v3 = v1  with the gaps fixed.</p>",
              "rawMarkdown": "v3 = v1  with the gaps fixed."
            },
            {
              "id": 3415192,
              "postDate": "2026-02-28T12:03:12.500Z",
              "content": "<p>Not quite because some transliterations have been truncated between v1 and v3</p>",
              "rawMarkdown": "Not quite because some transliterations have been truncated between v1 and v3",
              "votes": 2
            }
          ]
        }
      ]
    },
    {
      "id": 3414173,
      "postDate": "2026-02-26T06:35:04.343Z",
      "content": "<pre><code>Remove from translations:\n\n    fem.\n    sing.\n    pl.\n    plural\n    (?)\n    any stray marks you find (e.g., .., ?, x, xx, &lt;&lt; &gt;&gt;, &lt; &gt; except around &lt;gap&gt; of course)\n</code></pre>\n<p>You recommend removing these words but would these words be included in the test set?</p>",
      "rawMarkdown": "```\nRemove from translations:\n\n    fem.\n    sing.\n    pl.\n    plural\n    (?)\n    any stray marks you find (e.g., .., ?, x, xx, << >>, < > except around <gap> of course)\n```\nYou recommend removing these words but would these words be included in the test set?",
      "votes": 2,
      "replies": [
        {
          "id": 3416494,
          "postDate": "2026-03-03T02:44:20.393Z",
          "content": "<p>no, they are not included in the test set</p>",
          "rawMarkdown": "no, they are not included in the test set"
        }
      ]
    },
    {
      "id": 3414058,
      "postDate": "2026-02-25T20:44:30.280Z",
      "content": "<p>The updated data is now live on the site. I will commence the rescore of existing submissions shortly.</p>\n<p>UPDATE 02/26/2026: The rescore is now complete. Please let us know if you have any questions or concerns.</p>",
      "rawMarkdown": "The updated data is now live on the site. I will commence the rescore of existing submissions shortly.\n\nUPDATE 02/26/2026: The rescore is now complete. Please let us know if you have any questions or concerns.",
      "votes": 2,
      "isPinned": true,
      "replies": [
        {
          "id": 3414132,
          "postDate": "2026-02-26T02:22:15.833Z",
          "content": "<p>rip scores</p>",
          "rawMarkdown": "rip scores",
          "votes": 2
        },
        {
          "id": 3414351,
          "postDate": "2026-02-26T15:34:44.203Z",
          "content": "<p>Our score on LB is 34.1 but the best scoring sub in our sub history seems to be 33.5. I don't understand. 😅</p>",
          "rawMarkdown": "Our score on LB is 34.1 but the best scoring sub in our sub history seems to be 33.5. I don't understand. 😅",
          "votes": 1,
          "replies": [
            {
              "id": 3414411,
              "postDate": "2026-02-26T17:56:38.327Z",
              "content": "<p>Hi <a href=\"https://www.kaggle.com/nlztrk\" target=\"_blank\">@nlztrk</a>,</p>\n<p>Something seems to have gone wrong with the LB update. We are investigating.</p>",
              "rawMarkdown": "Hi @nlztrk,\n\nSomething seems to have gone wrong with the LB update. We are investigating.",
              "votes": 4
            },
            {
              "id": 3414443,
              "postDate": "2026-02-26T19:48:53.087Z",
              "content": "<p>Also the scores of some  <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/code?competitionId=121150&amp;sortBy=scoreDescending&amp;excludeNonAccessedDatasources=true\" target=\"_blank\">public NBs</a> have not been updated</p>",
              "rawMarkdown": "Also the scores of some  [public NBs](https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/code?competitionId=121150&sortBy=scoreDescending&excludeNonAccessedDatasources=true) have not been updated"
            },
            {
              "id": 3414524,
              "postDate": "2026-02-27T03:00:14.150Z",
              "content": "<p>Hi, my LB rank changed again (40+-&gt;30+) after rerun, is there new update?</p>",
              "rawMarkdown": "Hi, my LB rank changed again (40+->30+) after rerun, is there new update?"
            },
            {
              "id": 3414608,
              "postDate": "2026-02-27T08:56:32.717Z",
              "rawMarkdown": "",
              "isDeleted": true
            },
            {
              "id": 3414882,
              "postDate": "2026-02-27T21:15:48.293Z",
              "content": "<p>Any update on this?</p>",
              "rawMarkdown": "Any update on this?",
              "votes": 1
            },
            {
              "id": 3416289,
              "postDate": "2026-03-02T15:03:16.157Z",
              "content": "<p>I recalculated the leaderboard based on the rescore. It should be up to date now. Are you still seeing out of date scores?</p>",
              "rawMarkdown": "I recalculated the leaderboard based on the rescore. It should be up to date now. Are you still seeing out of date scores?"
            }
          ]
        },
        {
          "id": 3414535,
          "postDate": "2026-02-27T04:14:59.490Z",
          "content": "<p>transliteration:2-\ntranslation: 2+ or 2-\ntranslation of test is 2- or 2+ or 2(+gap) ???\nplease let me know.thank you </p>",
          "rawMarkdown": "transliteration:2-<gap>\ntranslation: 2+<gap> or 2-<gap>\ntranslation of test is 2-<gap> or 2+<gap> or 2(+gap) ???\nplease let me know.thank you "
        }
      ]
    },
    {
      "id": 3424553,
      "postDate": "2026-03-19T19:37:22.960Z",
      "content": "<p>Cool ,its intresting </p>",
      "rawMarkdown": "Cool ,its intresting ",
      "votes": -1
    },
    {
      "id": 3421313,
      "postDate": "2026-03-15T07:15:22.900Z",
      "content": "<p>Sorry, I still don't understand what a \"meaningful\" question mark and exclamation point are. If they're enclosed in parentheses with words like: (and?), should I change that to -&gt; (and) or leave them as is?</p>",
      "rawMarkdown": "Sorry, I still don't understand what a \"meaningful\" question mark and exclamation point are. If they're enclosed in parentheses with words like: (and?), should I change that to -> (and) or leave them as is?"
    },
    {
      "id": 3415409,
      "postDate": "2026-02-28T21:37:13.477Z",
      "content": "<p>1 / 12 (shekel) —&gt; 15 grains</p>\n<p>what should we do with 1/4, 1/6, and 1/8</p>",
      "rawMarkdown": "1 / 12 (shekel) —> 15 grains\n\nwhat should we do with 1/4, 1/6, and 1/8",
      "replies": [
        {
          "id": 3416498,
          "postDate": "2026-03-03T02:48:37.137Z",
          "content": "<p>Those fractions are counted in terms of shekels:\n¼ (shekel)\n⅙ (shekel)</p>",
          "rawMarkdown": "Those fractions are counted in terms of shekels:\n¼ (shekel)\n⅙ (shekel)",
          "votes": 1
        }
      ]
    },
    {
      "id": 3415296,
      "postDate": "2026-02-28T16:18:10.413Z",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a>  hi, just to be 100% clarified, is ḫ gonna be a valid character in the test set? </p>\n<p>Asking this because in the \"Examples of the Optimal Outcomes for Train\" you mentioned, the proper noun \"Tuwar-a<strong>ḫ</strong>šu\" is displayed as \"Tuwar-a<strong>h</strong>šu\" </p>\n<p>thanks!</p>",
      "rawMarkdown": "@deeppast  hi, just to be 100% clarified, is ḫ gonna be a valid character in the test set? \n\nAsking this because in the \"Examples of the Optimal Outcomes for Train\" you mentioned, the proper noun \"Tuwar-a**ḫ**šu\" is displayed as \"Tuwar-a**h**šu\" \n\nthanks!",
      "replies": [
        {
          "id": 3416499,
          "postDate": "2026-03-03T02:49:32.700Z",
          "content": "<p>yes, neither <code>Ḫ</code> nor <code>ḫ</code> appear in the test set.</p>",
          "rawMarkdown": "yes, neither `Ḫ` nor `ḫ` appear in the test set."
        }
      ]
    },
    {
      "id": 3414239,
      "postDate": "2026-02-26T09:59:06.320Z",
      "content": "<p>Hi! \nCan you clarify these examples in translation:</p>\n<pre><code>I paid &lt;gap&gt; ⅓ mina (silver) to Ennānum, son of Amriya.\nTuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.\n</code></pre>\n<p>Why are we using curved parenthesis? I looked at train.csv on datasets page and there are no examples of \"(silver)\". Or examples of parenthesis usage.</p>",
      "rawMarkdown": "Hi! \nCan you clarify these examples in translation:\n\n```python\nI paid <gap> ⅓ mina (silver) to Ennānum, son of Amriya.\nTuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.\n```\n\nWhy are we using curved parenthesis? I looked at train.csv on datasets page and there are no examples of \"(silver)\". Or examples of parenthesis usage.",
      "replies": [
        {
          "id": 3414462,
          "postDate": "2026-02-26T20:40:06.403Z",
          "content": "<p>The parentheses are used by translators to fill in words that are missing in the Akkadian transliteration, but which provides sometimes necessary context.</p>",
          "rawMarkdown": "The parentheses are used by translators to fill in words that are missing in the Akkadian transliteration, but which provides sometimes necessary context.",
          "votes": 2
        }
      ]
    },
    {
      "id": 3414137,
      "postDate": "2026-02-26T03:01:15.023Z",
      "content": "<p>Thanks for the update! </p>\n<p>May I ask about a specific case: 33773ec0-e74f-41bf-b985-f3e35b0f26c9</p>\n<pre><code>From what Amur-Aššur (owes?)\n</code></pre>\n<p>will such case of <code>(some word?)</code>  appear in the test set? </p>",
      "rawMarkdown": "Thanks for the update! \n\nMay I ask about a specific case: 33773ec0-e74f-41bf-b985-f3e35b0f26c9\n\n```plaintext\nFrom what Amur-Aššur (owes?)\n```\nwill such case of `(some word?)`  appear in the test set? ",
      "replies": [
        {
          "id": 3414460,
          "postDate": "2026-02-26T20:38:24.687Z",
          "content": "<p>yes, unfortunately, that does exist.</p>",
          "rawMarkdown": "yes, unfortunately, that does exist."
        }
      ]
    },
    {
      "id": 3413999,
      "postDate": "2026-02-25T17:52:29.277Z",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a> for your ongoing efforts. What does this exactly mean?</p>\n<blockquote>\n  <p>PN —&gt; &lt; gap &gt;</p>\n</blockquote>\n<p>Does this mean that every personal name to be found must be replaced by a gap tag? Or just the literal <code>PN</code>?</p>\n<p>Also, </p>\n<blockquote>\n  <p>Shortening of long floats to four places after the decimal</p>\n</blockquote>\n<p>I am a bit confused by this since later you suggest converting them to unicode fractions. Does the hidden test set (both translations and transliterations) contain floating point numbers or only unicode fractions?</p>\n<p>Thanks in advance </p>",
      "rawMarkdown": "Thanks @deeppast for your ongoing efforts. What does this exactly mean?\n> PN —> < gap >\n\nDoes this mean that every personal name to be found must be replaced by a gap tag? Or just the literal `PN`?\n\nAlso, \n> Shortening of long floats to four places after the decimal\n\nI am a bit confused by this since later you suggest converting them to unicode fractions. Does the hidden test set (both translations and transliterations) contain floating point numbers or only unicode fractions?\n\nThanks in advance ",
      "replies": [
        {
          "id": 3414000,
          "postDate": "2026-02-25T17:59:16.453Z",
          "content": "<p>That's a literal <code>PN</code> token, there are some of these in Veenhof's translations from AKT 8.</p>\n<p>I already shortened the floats, so that the conversion to fractions will be easier for you, if you choose to do so. As seen in the example at the end, the test contains only unicode fractions (no decimals at all).</p>",
          "rawMarkdown": "That's a literal `PN` token, there are some of these in Veenhof's translations from AKT 8.\n\nI already shortened the floats, so that the conversion to fractions will be easier for you, if you choose to do so. As seen in the example at the end, the test contains only unicode fractions (no decimals at all).",
          "votes": 3,
          "replies": [
            {
              "id": 3414003,
              "postDate": "2026-02-25T18:05:32.347Z",
              "content": "<p>Perfect, thanks a lot!</p>",
              "rawMarkdown": "Perfect, thanks a lot!"
            }
          ]
        }
      ]
    },
    {
      "id": 3414094,
      "postDate": "2026-02-25T22:55:42.473Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3414072,
      "author_name": "MPWARE",
      "author_url": "",
      "post_date": "2026-02-25T21:44:35.927000",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a> Thanks for the update. Some clarification to be 101% sure:</p>\n<p><strong>About double quotes:</strong></p>\n<blockquote>\n  <p>Do not remove from translations (as these are in the test too):</p>\n  <p>quotation marks \" \"</p>\n</blockquote>\n<p>In the previous update:</p>\n<blockquote>\n  <p>quotations \" “= removed</p>\n</blockquote>\n<p>The final rule is keep all quotation marks, correct? And for curly double quotes, we have to convert them to plain double quotes or not?</p>\n<p><strong>About  parentheses:</strong></p>\n<p>In the previous update:</p>\n<blockquote>\n  <p>parentheses ( )= removed</p>\n</blockquote>\n<p>Now, we need to keep parenthesis as for the examples outcomes you've provided, right?</p>\n<p>For instance: <code>0faa50f7-b86c-466c-a8ab-3a6f48fcb00a</code></p>\n<p>\"at 52.5 grains (of silver) each, 0.5 shekel\" we must keep the parenthesis around <code>of silver</code>, right?</p>\n<p>We must not do: \"at 52.5 grains of silver each, 0.5 shekel\"</p>\n<p><strong>About replacements:</strong></p>\n<blockquote>\n  <p>-tax —&gt; šadduātum tax</p>\n  <p>-gold —&gt; pašallum gold</p>\n  <p>textiles —&gt; kutānum textiles (are you sure it's not -textiles —&gt; kutānum textiles )?</p>\n</blockquote>\n<p>And it's <code>textiles</code> and <code>-tax</code> and <code>-gold</code> surrounded by spaces? \nIf not then in <code>198e428d-f51b-40d1-96d8-aee4bfa60d8d</code>:</p>\n<p>\"of Ennam-Suen, 8 textiles as <strong>import-tax,</strong> 10 textiles are pre-emption\" </p>\n<p>would become: </p>\n<p>\"of Ennam-Suen, 8 kutānum textiles as <strong>importšadduātum tax</strong>, 10 kutānum textiles are pre-emption\". </p>\n<p>and it 229be03b-c772-4a6c-8792-c3f80948a97d:</p>\n<p>\"received 28 kutānu-textiles and 1 black donkey\"</p>\n<p>would become: </p>\n<p>\"received 28 <strong>kutānu-kutānum textiles</strong> and 1 black donkey\"</p>\n<p>Which would look weird, please confirm.</p>\n<p><strong>About <code>&lt;gap&gt;</code>:</strong></p>\n<p>In previous update:</p>\n<blockquote>\n  <p>added space around gap in translations (not transliterations)</p>\n</blockquote>\n<p>Is it still true? Do we need to add space around <code>&lt;gap&gt;</code> except if glued with a dash?</p>",
      "votes": 12,
      "replies": [
        {
          "id": 3414734,
          "author_name": "Oleg Yaroshevskiy",
          "author_url": "",
          "post_date": "2026-02-27T14:50:56.837000",
          "content": "<p>the questions about quotes and parentheses are extremely important</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 3414893,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-27T21:59:35.493000",
          "content": "<p>Yes, keep parentheses and quotations should be dead quotes, no curl, same for apostrophes / scare quotes.</p>\n<blockquote>\n  <p>Do not remove from translations (as these are in the test too):\n  quotation marks <code>\" \"</code>\n  apostrophes <code>'</code>\n  meaningful question marks <code>?</code> or exclamation marks <code>!</code></p>\n</blockquote>\n<p>As for the words which begin with <code>-</code>: words which appear with an initial hyphen have a space before them, rather than those words which are joined to another word with a hyphen. These suggestions are meant to provide some context for the missing words, but the only way to know for sure is to check in the PDFs, which is why they were not replaced in v3. These missing parts were removed in the OARE database, and should be replaced in an ideal situation by checking the PDFs.</p>\n<ul>\n<li><code>-gold</code> —&gt; <code>pašallum gold</code></li>\n<li><code>-tax</code> —&gt; <code>šadduātum tax</code></li>\n<li><code>-textiles</code> —&gt; <code>kutānum textiles</code></li>\n</ul>",
          "votes": 4,
          "replies": []
        }
      ]
    },
    {
      "id": 3416050,
      "author_name": "Ricardo Pérez",
      "author_url": "",
      "post_date": "2026-03-02T00:10:52.533000",
      "content": "<p>Thanks for the comprehensive update and all the clarifications in the comments. After applying the v3 changes, a few things that helped on my end:</p>\n<p>Recovering truncated transliterations: Confirmed that ~10% are affected. Matching oare_id against published_texts.csv and cross-referencing with the AKT PDFs recovered most of them. Be careful, though,  some entries aren't just truncated; the translation doesn't match the transliteration at all (as MPWARE pointed out for AKT 8, 55), so it's worth doing a length-ratio sanity check before blindly patching.</p>\n<p>The -textiles / -gold / -tax replacements: Based on the host's clarification, these only apply when preceded by a space (i.e., -textiles to kutānum textiles), not in the middle of a word, like import-tax or kutānu-textiles. A naive <em>str.replace</em> will break things, regex with word boundaries or explicit space matching is safer here.</p>\n<p>Straight quotes: Curly quotes to straight quotes for both \" and '. Small thing, but easy to miss if your text editor or PDF extraction reintroduces curly ones.</p>\n<p>Still working through the fraction conversions and subscript normalization. Good luck to everyone in the final stretch!</p>",
      "votes": 7,
      "replies": []
    },
    {
      "id": 3414172,
      "author_name": "",
      "author_url": "",
      "post_date": "2026-02-26T06:35:02.263000",
      "content": "<p>I think you should update these data from the beginning.Some samples from train.csv were aligned by hand. To revise these one more is really a challenging work.</p>",
      "votes": 7,
      "replies": [
        {
          "id": 3414894,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-27T21:59:51.867000",
          "content": "<p>There will be no further updates, I'm sorry.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 3414169,
      "author_name": "Yicong XIAO",
      "author_url": "",
      "post_date": "2026-02-26T06:28:22.013000",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a>  hihi, </p>\n<p>Regarding the unit conversions that you mentioned below, I suspect that the right-hand-side of the first two rows should be interchanged</p>\n<pre><code>1 / 12 (shekel) —&gt; ⅔ shekel 15 grains\n5 / 12 shekel —&gt; 15 grains\n5 11 / 12 shekels —&gt; 6 shekels less 15 grains\n7 / 12 shekel —&gt; ½ shekel 15 grains\n</code></pre>\n<p>Is it that <code>1/12 shekel == 15 grains</code> ?</p>\n<p>if so, then the conversions should instead be </p>\n<pre><code>1 / 12 (shekel) —&gt; 15 grains\n5 / 12 shekel —&gt; ⅔ shekel 15 grains\n5 11 / 12 shekels —&gt; 6 shekels less 15 grains\n7 / 12 shekel —&gt; ½ shekel 15 grains\n</code></pre>\n<p>is it?</p>",
      "votes": 6,
      "replies": [
        {
          "id": 3414891,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-27T21:43:03.437000",
          "content": "<p>close, 5 / 12 --&gt; ⅓ shekel 15 grains</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 3414997,
          "author_name": "esprit",
          "author_url": "",
          "post_date": "2026-02-28T03:33:12.203000",
          "content": "<p>Isn't it <code>5 / 12 shekel —&gt; ⅓ shekel 15 grains</code> instead of <code>5 / 12 shekel —&gt; ⅔ shekel 15 grains</code>?</p>\n<p>There are probably many similar mistakes in the test data.</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3415162,
              "author_name": "Yicong XIAO",
              "author_url": "",
              "post_date": "2026-02-28T10:14:05.900000",
              "content": "<p>oops nice catch you are right</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414139,
      "author_name": "epikt",
      "author_url": "",
      "post_date": "2026-02-26T03:27:06.307000",
      "content": "<p>I am a bit confused about the table that follows 'Results of these changes: no more <code>&lt;big_gap&gt;</code>'. That table still says <code>&lt;gap&gt; &lt;gap&gt;: 30</code>, <code>&lt;gap&gt; &lt;big_gap&gt;: 6</code>, etc. Does this mean that there's till 30 occurrences of <code>&lt;gap&gt; &lt;gap&gt;</code>, 6 occurrences of <code>&lt;gap&gt; &lt;big_gap&gt;</code>, etc. in the test set?</p>\n<p>I thought I had understood that all of the <code>&lt;big_gap&gt;</code> had been replaced by <code>&lt;gap&gt;</code> and neighboring gaps had been merged together, but I'm less certain I understand right now.</p>\n<p>Edit: it seems that some time since this post <code>Results of these changes: no more &lt;big_gap&gt;</code> was changed to <code>Results of these changes: no more &lt;big_gap&gt; and no more duplicates for &lt;gap&gt; Before the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single &lt;gap&gt;.</code> so it's now clearer. Thank you for the clarification!</p>",
      "votes": 6,
      "replies": [
        {
          "id": 3414146,
          "author_name": "hongan",
          "author_url": "",
          "post_date": "2026-02-26T03:47:32.113000",
          "content": "<p>In this very post, i can't find any info about merging gap…</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3414148,
              "author_name": "epikt",
              "author_url": "",
              "post_date": "2026-02-26T03:53:39.457000",
              "content": "<p>That's what had been mentioned in last week's update:</p>\n<pre><code>Gaps, damage markers, and parallel alignment [update: 2/18/26]\n\nAnother recurring source of confusion concerns damaged text and gap markers in the data:\n\n    x represents a single broken sign,\n    sequences like x x x x or ... represent a larger lacuna.\n\nFor modeling purposes we reduced all breaks to a single marker: &lt;gap&gt;\n\n we removed the tag for &lt;big_gap&gt; from the train and test (and other transliterations). We also deduplicated instances multiple sequential gaps (e.g. &lt;gap&gt; &lt;gap, &lt;gap&gt;-&lt;gap&gt;, &lt;gap&gt; &lt;gap&gt;, &lt;gap&gt;. &lt;gap&gt;, etc.\n</code></pre>\n<p>Source: <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209\" target=\"_blank\">https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/665209</a></p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3414149,
              "author_name": "hongan",
              "author_url": "",
              "post_date": "2026-02-26T03:58:27.463000",
              "content": "<p>thanks, i'm aware, just thought this post is a comprehensive summary of all the changes, which apparently is not…</p>",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3418958,
      "author_name": "Geremie Yeo",
      "author_url": "",
      "post_date": "2026-03-09T15:51:14.303000",
      "content": "<p>Sorry if I am being dumb here</p>\n<pre><code>The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:\n</code></pre>\n<p>why is it “no guarantee”? IMO these are deterministic postprocessing, so if the test set conforms to this, the score will 100% improve. Does this mean only some texts were changed in this manner while the others are left as it is?</p>",
      "votes": 3,
      "replies": [
        {
          "id": 3419052,
          "author_name": "tennogh",
          "author_url": "",
          "post_date": "2026-03-09T19:29:08.723000",
          "content": "<p>Others have pointed that some of those recommendations decrease LB score. I'm guessing either the test set is inconsistent, or it doens't respect those guidelines at all.</p>",
          "votes": 3,
          "replies": []
        }
      ]
    },
    {
      "id": 3415113,
      "author_name": "FX-6300",
      "author_url": "",
      "post_date": "2026-02-28T07:33:26.507000",
      "content": "<p>It looks like the “Change Roman numerals to integer numbers for months” table is misaligned/broken: Month 2 is labeled as Roman III, and III appears twice. </p>\n<p>According to Kouwenberg (Introduction to Old Assyrian, p. 182), the Old Assyrian calendar starts with Month 1 = I (Bēlat ekallem/Bēltekallem), followed by Month 2 = II (ša sarrātim) and Month 3 = III (ša kēnātim).\nCould you add Month 1 and shift/correct the Roman numerals accordingly?</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 3414479,
      "author_name": "Steve Roberts",
      "author_url": "",
      "post_date": "2026-02-26T21:40:52.667000",
      "content": "<p>The new update seems to have truncated the longer transliteration strings. The previous training data had the following long training entries:</p>\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>592</td>\n      <td>746</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>338</td>\n      <td>588</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>343</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>408</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>357</td>\n      <td>626</td>\n    </tr>\n    <tr>\n      <td>1260</td>\n      <td>c97bb594-a5a1-4674-9496-48496e91c2ee</td>\n      <td>29</td>\n      <td>602</td>\n    </tr>\n    <tr>\n      <td>1378</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>379</td>\n      <td>603</td>\n    </tr>\n  </tbody>\n</table>\n<p><br>\n<strong>The new data has truncated these to be:</strong></p>\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>138</td>\n      <td>739</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>107</td>\n      <td>587</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>114</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>108</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>107</td>\n      <td>625</td>\n    </tr>\n    <tr>\n      <td>1377</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>120</td>\n      <td>585</td>\n    </tr>\n  </tbody>\n</table>\n<p><br>\nI'm not sure if this applies to the other rows too, but it certainly seems to have lost a lot of data.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 3414484,
          "author_name": "MPWARE",
          "author_url": "",
          "post_date": "2026-02-26T22:03:52.923000",
          "content": "<p>Around 10% of transliterations are truncated. We've to fix that ourself according to the instructions.</p>",
          "votes": 1,
          "replies": [
            {
              "id": 3414560,
              "author_name": "steubk",
              "author_url": "",
              "post_date": "2026-02-27T05:52:55.683000",
              "content": "<p>which instructions?</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3414580,
              "author_name": "MPWARE",
              "author_url": "",
              "post_date": "2026-02-27T07:36:24.560000",
              "content": "<blockquote>\n  <p>Some of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique OARE_Text_ID with the equivalent OARE IDs in the published_texts.csv dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (<a href=\"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)\" target=\"_blank\">https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data)</a>.</p>\n</blockquote>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3414593,
              "author_name": "Steve Roberts",
              "author_url": "",
              "post_date": "2026-02-27T08:18:10.987000",
              "content": "<p>It's a bit bad though that data that was there, and which I spent a large amount of time working out how to split, has now suddenly disappeared from the data set. </p>",
              "votes": 5,
              "replies": []
            }
          ]
        },
        {
          "id": 3415110,
          "author_name": "Yicong XIAO",
          "author_url": "",
          "post_date": "2026-02-28T07:26:44.910000",
          "content": "<p>A large number of data where the ratio between the lengths of transliteration and translation seem weird have wrong/truncated transliteration/translation, e.g. 8376cbda-b423-42d4-abb5-188d04896392. Can follow the instruction to find the raw pdf and amend these data case by case.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3416596,
      "author_name": "Kh0a",
      "author_url": "",
      "post_date": "2026-03-03T10:48:47.297000",
      "content": "<p>question about (n broken lines), do the lines between \"Nimar-Istar.\" and \"When he returns from Mamma he will bring\nit to me\" counts as n broken lines?</p>\n<p>If I understand correctly, the translation should be:</p>\n<p>I furthermore gave 1 mina of good, native copper for an allu’āru-container of sweet wine from Mamma to Puzur-Amurrum son of Nimar-Ištar. &lt;gap&gt; When he returns from Mamma he will bring it to me. (This was) apart from the 4 allu’āru-containers, the proceeds from the silver and washed copper that they have received (previously). Witnessed by Šu-Bēlum son of Kuzizia.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21691005%2F5a2fcafea533880783e26906cb8cdffd%2FScreenshot%202026-03-03%20174635.png?generation=1772534814707795&amp;alt=media\" alt=\"\"></p>",
      "votes": 1,
      "replies": [
        {
          "id": 3419308,
          "author_name": "tennogh",
          "author_url": "",
          "post_date": "2026-03-10T12:22:04.547000",
          "content": "<p>I think in this case there is no gap in the translation, it just appears that way because of the alignment.</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 3421590,
          "author_name": "riberoc",
          "author_url": "",
          "post_date": "2026-03-15T22:05:18.830000",
          "content": "<p>Where are you getting these translations from?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3421594,
              "author_name": "tennogh",
              "author_url": "",
              "post_date": "2026-03-15T22:24:23.353000",
              "content": "<p>It's the PDFs from the second link here: <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/668485\" target=\"_blank\">https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/discussion/668485</a></p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414925,
      "author_name": "tennogh",
      "author_url": "",
      "post_date": "2026-02-27T23:53:45.343000",
      "content": "<p>Regarding the terms \"pašallum gold\", \"šadduātum tax\", \"kutānum textiles\", those appear in different forms in the texts. Are those the forms that are expected from the test data (e.g. \"kutanum-textile\", \"kutanu textile\" -&gt; \"kutānum textiles\")?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3414939,
          "author_name": "耶✌",
          "author_url": "",
          "post_date": "2026-02-28T00:55:34.600000",
          "content": "<p>I don’t think that’s the case. </p>\n<p>First, words like pašallum, šadduātum, and kutānum do not exist in train.csv at all. </p>\n<p>Second, submissions that do not include words such as pašallum, šadduātum, and kutānum have achieved higher scores.</p>\n<p>The host only seemed to suggest that we do this. :)</p>",
          "votes": 4,
          "replies": [
            {
              "id": 3414946,
              "author_name": "",
              "author_url": "",
              "post_date": "2026-02-28T01:36:26.513000",
              "content": "<p>when i drop the '[', ['] in the train, my scores would drop.  Normally, if dropping the \"[\" and \"]\" , the scores would have been improved. I don't know why it is like this.</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3415026,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2026-02-28T04:35:10.683000",
              "content": "<p>In your transliterations or translations? </p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414179,
      "author_name": "steubk",
      "author_url": "",
      "post_date": "2026-02-26T06:44:41.323000",
      "content": "<p>just for the record, <strong>this update does not include all the changes made in the previous update</strong><br>\nFor example unicode subscript numbers are now present in the train transliteration:<br>\n<code>0064939c-59b9-4448-a63d-34612af0a1b5 -&gt; 1 TÚG ša qá-tim i-tur₄-DINGIR il₅-qé ...</code></p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3413998,
      "author_name": "Anil Ozturk",
      "author_url": "",
      "post_date": "2026-02-25T17:50:26.173000",
      "content": "<p>Thanks, these are some new changes. Do we expect another test set update?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3414001,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-25T18:01:19.850000",
          "content": "<p>Yes, both test and train were updated (final update 2/26/26).</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3415275,
      "author_name": "MPWARE",
      "author_url": "",
      "post_date": "2026-02-28T15:21:30.847000",
      "content": "<p>I'm trying to identify the mapping between oare_id in train.csv and the AKT PDF in order to complete broken sections, I've found 1561 matches:</p>\n<ul>\n<li>AKT 5 = 77 oare_id </li>\n<li>AKT 6a = 304 oare_id</li>\n<li>AKT 6b = 222 oare_id</li>\n<li>AKT 6c = 201 oare_id</li>\n<li>AKT 6d = 139 oare_id</li>\n<li>AKT 6e = 255 oare_id</li>\n<li>AKT 8 = 363 oare_id (e.g. <code>0123a9b9-e81e-4d7a-a79b-10e7c0aacbb9</code> Kt 91/k 471, page 213)</li>\n</ul>\n<p>Someone with similar results? <a href=\"https://www.kaggle.com/jackvd\" target=\"_blank\">@jackvd</a> You was saying around 400 items was broken?</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3415360,
          "author_name": "Angantyr",
          "author_url": "",
          "post_date": "2026-02-28T18:50:07.417000",
          "content": "<p>I have similar results with regard to AKT 5 and AKT 6a. I still have to verify a few ids to be sure but when I do I can post it in the updated <code>train.csv</code> dataset to make everyone lives a bit easier.</p>\n<p>Edit: The <a href=\"https://www.kaggle.com/datasets/angantyr/updated-train-w-publication-source-96-7-cov/data\" target=\"_blank\">dataset with sources included</a> should be up and running.</p>\n<table>\n<thead>\n<tr>\n<th>Publication</th>\n<th>page</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>AKT 5</td>\n<td>76</td>\n</tr>\n<tr>\n<td>AKT 6a</td>\n<td>295</td>\n</tr>\n<tr>\n<td>AKT 6b</td>\n<td>218</td>\n</tr>\n<tr>\n<td>AKT 6c</td>\n<td>196</td>\n</tr>\n<tr>\n<td>AKT 6d</td>\n<td>137</td>\n</tr>\n<tr>\n<td>AKT 6e</td>\n<td>233</td>\n</tr>\n<tr>\n<td>AKT 8</td>\n<td>355</td>\n</tr>\n</tbody>\n</table>",
          "votes": 6,
          "replies": [
            {
              "id": 3415774,
              "author_name": "MPWARE",
              "author_url": "",
              "post_date": "2026-03-01T09:50:59.623000",
              "content": "<p>I've updated my list:</p>\n<pre><code>pub_pd = pd.read_csv(\"data/published_texts.csv\")\n\ntrain_df = pd.read_csv(\"data/train.csv\")\n\nakt_pd = pub_pd[(pub_pd['oare_id'].isin(train_df[\"oare_id\"].unique()))][[\"oare_id\", \"label\", \"excavation_no\", \"transliteration\"]]\nakt_pd[\"pdf\"] = akt_pd[\"label\"].str.extract(\n    r\"\\bAKT\\s+(\\d+[a-zA-Z]?)\",\n    expand=False\n)\n</code></pre>\n<pre><code>pdf\n5      77\n6a    304\n6b    222\n6c    201\n6d    139\n6e    255\n8     363\n</code></pre>\n<p>Some rows in train.csv are more than broken, the translation just does not match the transliteration at all. For instance: AKT8, 55. Kt 91/k 304 (1-161-91) oare_id: <code>5f088d12-ed99-434a-a113-65deab7e1426</code></p>\n<p><strong>In the PDF:</strong>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F11e1483f1b5212cfa6310549bc4f6a4e%2Fk304.png?generation=1772371430299359&amp;alt=media\" alt=\"\"></p>\n<p><strong>In train.csv:</strong> whatever v1,v2 or v3\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2F838802a47acb10736ea3c509af8d3116%2Fk304-bad.png?generation=1772371460757455&amp;alt=media\" alt=\"\"></p>\n<p><strong>After RE-OCR:</strong>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2Fb94e2b6c02e584b71d68d4246257ccb9%2Fk304-fixed.png?generation=1772372008286642&amp;alt=media\" alt=\"\"></p>\n<p>oare_ids from AKT 6e are one of the most broken.</p>\n<p>This competition is about normalization and cleaning.</p>",
              "votes": 5,
              "replies": []
            },
            {
              "id": 3416194,
              "author_name": "Angantyr",
              "author_url": "",
              "post_date": "2026-03-02T09:25:40.107000",
              "content": "<p>I wonder for how many of such cases we could employ a set match filtering, e.g., 'a-na'/'from', 'ku.babbar'/'silver', etc.</p>\n<p>It would not be a guarantee but a first stage flag to have a manual check and there are many words that have a 1:1 equivalent with English.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414364,
      "author_name": "Oleg Yaroshevskiy",
      "author_url": "",
      "post_date": "2026-02-26T16:02:45.813000",
      "content": "<p>Can't describe an amount of confusion when comparing a new train dropped last week and a new train from yesterday and now I'm trying do understand how to merge those two</p>\n<p>upd: sorry, my humble verdict is that these updates last week made things worse…</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3414370,
          "author_name": "Oleg Yaroshevskiy",
          "author_url": "",
          "post_date": "2026-02-26T16:14:06.147000",
          "content": "<p>how to understand this:</p>\n<pre><code>before: 31 kutānu-textiles of Iddin-Suen, 30 kutānu-textiles of Ah-šalim, 46 kutānu-textiles of hinnāya and Uṣurānum, 21 kutānu-textiles of Aššur-rēī, 36 kutānu-textiles of Šu-Ištar, 28 kutānu-textiles of Ennam-Suen, 24 kutānu-textiles of Lā-qēp, 18 kutānu-textiles of the merchant, 7 kutānu-textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\nafter:  31 -textiles of Iddin-Suen, 30 -textiles of Ah-šalim, 46 -textiles of hinnāya and Uṣurānum, 21 -textiles of Aššur-rē'ī, 36 -textiles of Šu-Ištar, 28 -textiles of Ennam-Suen, 24 -textiles of Lā-qēp, 18 -textiles of the merchant, 7 -textiles without seals, 14 textiles import-tax, 17 textiles as pre-emption.\n</code></pre>",
          "votes": 0,
          "replies": [
            {
              "id": 3414465,
              "author_name": "Adam Anderson",
              "author_url": "",
              "post_date": "2026-02-26T20:43:13.670000",
              "content": "<p>Yes, you can see in the post above, I recommend making some small changes, egg. <code>-textiles</code> --&gt; <code>kutānum textiles</code>, but this should be checked with the AKT volumes in PDF to be certain. The interum data (v2) update did this, but it introduced new errors because of a simple search / replace. Those changes were rolled back in the lates update (v3), and instructions were provided how to deal with those (above).</p>",
              "votes": -1,
              "replies": []
            },
            {
              "id": 3414732,
              "author_name": "Oleg Yaroshevskiy",
              "author_url": "",
              "post_date": "2026-02-27T14:49:29.263000",
              "content": "<p>do you recommend to take V3 and manually update it based on the tutorial or to take V2 and just fix 's and quotes?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3414898,
              "author_name": "tennogh",
              "author_url": "",
              "post_date": "2026-02-27T22:22:59.053000",
              "content": "<p>Maybe even v1 is the least flawed, especially if you have already done the work to fix the gaps on it.</p>",
              "votes": 2,
              "replies": []
            },
            {
              "id": 3414927,
              "author_name": "Adam Anderson",
              "author_url": "",
              "post_date": "2026-02-27T23:58:08.460000",
              "content": "<p>v3 = v1  with the gaps fixed.</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3415192,
              "author_name": "tennogh",
              "author_url": "",
              "post_date": "2026-02-28T12:03:12.500000",
              "content": "<p>Not quite because some transliterations have been truncated between v1 and v3</p>",
              "votes": 2,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414173,
      "author_name": "rejk",
      "author_url": "",
      "post_date": "2026-02-26T06:35:04.343000",
      "content": "<pre><code>Remove from translations:\n\n    fem.\n    sing.\n    pl.\n    plural\n    (?)\n    any stray marks you find (e.g., .., ?, x, xx, &lt;&lt; &gt;&gt;, &lt; &gt; except around &lt;gap&gt; of course)\n</code></pre>\n<p>You recommend removing these words but would these words be included in the test set?</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3416494,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-03-03T02:44:20.393000",
          "content": "<p>no, they are not included in the test set</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3414058,
      "author_name": "Ryan Holbrook",
      "author_url": "",
      "post_date": "2026-02-25T20:44:30.280000",
      "content": "<p>The updated data is now live on the site. I will commence the rescore of existing submissions shortly.</p>\n<p>UPDATE 02/26/2026: The rescore is now complete. Please let us know if you have any questions or concerns.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 3414132,
          "author_name": "Jack",
          "author_url": "",
          "post_date": "2026-02-26T02:22:15.833000",
          "content": "<p>rip scores</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 3414351,
          "author_name": "Anil Ozturk",
          "author_url": "",
          "post_date": "2026-02-26T15:34:44.203000",
          "content": "<p>Our score on LB is 34.1 but the best scoring sub in our sub history seems to be 33.5. I don't understand. 😅</p>",
          "votes": 1,
          "replies": [
            {
              "id": 3414411,
              "author_name": "Ryan Holbrook",
              "author_url": "",
              "post_date": "2026-02-26T17:56:38.327000",
              "content": "<p>Hi <a href=\"https://www.kaggle.com/nlztrk\" target=\"_blank\">@nlztrk</a>,</p>\n<p>Something seems to have gone wrong with the LB update. We are investigating.</p>",
              "votes": 4,
              "replies": []
            },
            {
              "id": 3414443,
              "author_name": "Mattia Angeli",
              "author_url": "",
              "post_date": "2026-02-26T19:48:53.087000",
              "content": "<p>Also the scores of some  <a href=\"https://www.kaggle.com/competitions/deep-past-initiative-machine-translation/code?competitionId=121150&amp;sortBy=scoreDescending&amp;excludeNonAccessedDatasources=true\" target=\"_blank\">public NBs</a> have not been updated</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3414524,
              "author_name": "guoqin gu",
              "author_url": "",
              "post_date": "2026-02-27T03:00:14.150000",
              "content": "<p>Hi, my LB rank changed again (40+-&gt;30+) after rerun, is there new update?</p>",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3414608,
              "author_name": "",
              "author_url": "",
              "post_date": "2026-02-27T08:56:32.717000",
              "content": "",
              "votes": 0,
              "replies": []
            },
            {
              "id": 3414882,
              "author_name": "Jack",
              "author_url": "",
              "post_date": "2026-02-27T21:15:48.293000",
              "content": "<p>Any update on this?</p>",
              "votes": 1,
              "replies": []
            },
            {
              "id": 3416289,
              "author_name": "Ryan Holbrook",
              "author_url": "",
              "post_date": "2026-03-02T15:03:16.157000",
              "content": "<p>I recalculated the leaderboard based on the rescore. It should be up to date now. Are you still seeing out of date scores?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3414535,
          "author_name": "",
          "author_url": "",
          "post_date": "2026-02-27T04:14:59.490000",
          "content": "<p>transliteration:2-\ntranslation: 2+ or 2-\ntranslation of test is 2- or 2+ or 2(+gap) ???\nplease let me know.thank you </p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3424553,
      "author_name": "Muhammed Shafeeh",
      "author_url": "",
      "post_date": "2026-03-19T19:37:22.960000",
      "content": "<p>Cool ,its intresting </p>",
      "votes": -1,
      "replies": []
    },
    {
      "id": 3421313,
      "author_name": "Iurii Samusevich",
      "author_url": "",
      "post_date": "2026-03-15T07:15:22.900000",
      "content": "<p>Sorry, I still don't understand what a \"meaningful\" question mark and exclamation point are. If they're enclosed in parentheses with words like: (and?), should I change that to -&gt; (and) or leave them as is?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 3415409,
      "author_name": "MengYe",
      "author_url": "",
      "post_date": "2026-02-28T21:37:13.477000",
      "content": "<p>1 / 12 (shekel) —&gt; 15 grains</p>\n<p>what should we do with 1/4, 1/6, and 1/8</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3416498,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-03-03T02:48:37.137000",
          "content": "<p>Those fractions are counted in terms of shekels:\n¼ (shekel)\n⅙ (shekel)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 3415296,
      "author_name": "Yicong XIAO",
      "author_url": "",
      "post_date": "2026-02-28T16:18:10.413000",
      "content": "<p><a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a>  hi, just to be 100% clarified, is ḫ gonna be a valid character in the test set? </p>\n<p>Asking this because in the \"Examples of the Optimal Outcomes for Train\" you mentioned, the proper noun \"Tuwar-a<strong>ḫ</strong>šu\" is displayed as \"Tuwar-a<strong>h</strong>šu\" </p>\n<p>thanks!</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3416499,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-03-03T02:49:32.700000",
          "content": "<p>yes, neither <code>Ḫ</code> nor <code>ḫ</code> appear in the test set.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3414239,
      "author_name": "Sasha Turutin",
      "author_url": "",
      "post_date": "2026-02-26T09:59:06.320000",
      "content": "<p>Hi! \nCan you clarify these examples in translation:</p>\n<pre><code>I paid &lt;gap&gt; ⅓ mina (silver) to Ennānum, son of Amriya.\nTuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.\n</code></pre>\n<p>Why are we using curved parenthesis? I looked at train.csv on datasets page and there are no examples of \"(silver)\". Or examples of parenthesis usage.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 3414462,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-26T20:40:06.403000",
          "content": "<p>The parentheses are used by translators to fill in words that are missing in the Akkadian transliteration, but which provides sometimes necessary context.</p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 3414137,
      "author_name": "Yicong XIAO",
      "author_url": "",
      "post_date": "2026-02-26T03:01:15.023000",
      "content": "<p>Thanks for the update! </p>\n<p>May I ask about a specific case: 33773ec0-e74f-41bf-b985-f3e35b0f26c9</p>\n<pre><code>From what Amur-Aššur (owes?)\n</code></pre>\n<p>will such case of <code>(some word?)</code>  appear in the test set? </p>",
      "votes": 0,
      "replies": [
        {
          "id": 3414460,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-26T20:38:24.687000",
          "content": "<p>yes, unfortunately, that does exist.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 3413999,
      "author_name": "moth",
      "author_url": "",
      "post_date": "2026-02-25T17:52:29.277000",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/deeppast\" target=\"_blank\">@deeppast</a> for your ongoing efforts. What does this exactly mean?</p>\n<blockquote>\n  <p>PN —&gt; &lt; gap &gt;</p>\n</blockquote>\n<p>Does this mean that every personal name to be found must be replaced by a gap tag? Or just the literal <code>PN</code>?</p>\n<p>Also, </p>\n<blockquote>\n  <p>Shortening of long floats to four places after the decimal</p>\n</blockquote>\n<p>I am a bit confused by this since later you suggest converting them to unicode fractions. Does the hidden test set (both translations and transliterations) contain floating point numbers or only unicode fractions?</p>\n<p>Thanks in advance </p>",
      "votes": 0,
      "replies": [
        {
          "id": 3414000,
          "author_name": "Adam Anderson",
          "author_url": "",
          "post_date": "2026-02-25T17:59:16.453000",
          "content": "<p>That's a literal <code>PN</code> token, there are some of these in Veenhof's translations from AKT 8.</p>\n<p>I already shortened the floats, so that the conversion to fractions will be easier for you, if you choose to do so. As seen in the example at the end, the test contains only unicode fractions (no decimals at all).</p>",
          "votes": 3,
          "replies": [
            {
              "id": 3414003,
              "author_name": "moth",
              "author_url": "",
              "post_date": "2026-02-25T18:05:32.347000",
              "content": "<p>Perfect, thanks a lot!</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    },
    {
      "id": 3414094,
      "author_name": "",
      "author_url": "",
      "post_date": "2026-02-25T22:55:42.473000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3413989": "If only the ancients who wrote these texts could imagine how difficult it would be to train computers to translate their work! Working with data from 1950-1750 BCE is not for the faint of heart. From 1927 CE until today, modern scholars have been working with the editing and publication of these ca. 8,000 documents in many different languages, and despite (or because of) almost 100 years of scholarly scrutiny, the datasets are more complicated and in a more convoluted format, than what the ancients originally wrote. These editions were migrated into a database, which became the foundation of the data we used for the training and supplemental data for this challenge. Unbeknownst to them, we are now re-purposing their editions for training ML models, and the messiness of the data is palpable for all to see. That said, this is the first benchmark for a significant ML application with the Old Assyrian data, so with great struggle comes greater pioneering outcomes.\n\nThe purpose of this final update is to level the number of markers for gaps from two to one. The hope is that by making some small, but significant adjustments, this will improve the scores for everyone. What we refrained from doing was making any additional changes which could introduce new inconsistencies if not done uniformly. So for the final month of the competition, we will include a clear and concise list of recommendations for the training dataset for alignment with the test data.\n\n**Gap Change: Replaced all the following with a single `<gap>`**\n- `x` —> `<gap>`\n- `[x]` —> `<gap>`\n- `…` —> `<gap>`\n- `(break)` —> `<gap>`\n- `(large break)` —> `<gap>`\n- `(n broken lines)` —> `<gap>`\n- `<gap>` `<gap>` —> `<gap>`\n\n**Results of these changes: no more `<big_gap>` and no more duplicates for `<gap>`** \nBefore the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single `<gap>`.\n\n| Gap | Count |\n|-----|-------|\n|`<gap> <gap>`|30|\n|`-<gap> <gap>`|9|\n|`-<gap>-<gap>`|1|\n|`<gap> <gap>-`|15|\n|`<gap> <big_gap>`|6|\n|`<gap> <gap> <gap>`|2|\n|`<gap> <big_gap> <gap>`|1|\n|`-<big_gap> <gap>-`|4|\n|`-<big_gap> <big_gap>`|13|\n|`<big_gap> <big_gap>`|15|\n|`<big_gap> <big_gap>-`|12|\n|`<big_gap> <gap> <big_gap>`|2|\n|`<big_gap> <big_gap> <big_gap> <big_gap>-`|3|\n|`<big_gap> <big_gap> <big_gap> <gap> <gap>`|1|\n|`-<big_gap> <big_gap> <gap> <big_gap> <big_gap>-`|1|\n\n\n**Alignment of Determinatives to match test data:**\n- `(d)` —> `{d}`\n- `(ki)` —> `{ki}`\n- `(TÚG)` —> `TÚG`\n\n**Shortening of long floats to four places after the decimal**\n- `1.3333300000000001` —> `1.3333`\n- `2.6666600000000003` —> `2.6666`\n\n**The rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:**\n\nSome of the transliteration text in the training data is missing, which is unfortunate, but fixable. You can find what is missing from the training data by matching the unique `OARE_Text_ID` with the equivalent OARE IDs in the `published_texts.csv` dataset. Do this by matching the unique IDs in the training data to the OARE IDs and find the publication (e.g. AKT 8, 130 = AKT volume 8, text number 130). The PDFs have also been provided, if you find there are missing elements in the translations (https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf/data).\n\n__Here are some recommended options for changes to the training data:__\n\n**Remove from translations:**\n- `fem.`\n- `sing.`\n- `pl.`\n- `plural`\n- `(?)`\n- any stray marks you find (e.g., `..`, `?`, `x`, `xx`, `<< >>`, `< >` except around `<gap>` of course)\n- some of the translations equivocate two optional translations using `/`, it might be better to choose one or the other options provided, rather than including both with a `/` (e.g. \"you / she brought\" —> \"you brought\" ).\n\n**Do not remove from translations (as these are in the test too):**\n- quotation marks `\" \"`\n- apostrophes `'`\n- meaningful question marks `?` or exclamation marks `!`\n\n**Replace in translations:**\n- `PN` —> `<gap>`\n- ` -gold` —> `pašallum gold`\n- ` -tax` —> `šadduātum tax`\n- ` -textiles` —> `kutānum textiles`\n- `1 / 12 (shekel)` —> `15 grains`\n- `5 / 12 shekel` —> `⅓ shekel 15 grains`\n- `5 11 / 12 shekels` —> `6 shekels less 15 grains`\n- `7 / 12 shekel` —> `½ shekel 15 grains`\n\n**Decimals to Fractions:**\n- `0.5`  —> `½`\n- `0.25` —> `¼`\n- `0.3333` —> `⅓`\n- `0.8333` —> `⅚`\n- `0.625` —> `⅝`\n- `0.6666` —> `⅔`\n- `0.75` —>  `¾`\n- `0.1666` —> `⅙`\n\n**Change Roman numerals to integer numbers for months: e.g., month V —> month 5**\n\n| Month | Roman | MN | forms | AKA |\n|--------|--------|-----|-------|------|\n|Month 1|\tI |\tBēlat ekallem | be-el-tí-É.GAL-lim; be-el-té-kà-lim |Bēlat-ekallim|\n|Month 2|\tII |\tša-sarratim |\tša sá-ra-tim|\t|\n|Month 3|\tIII |\tKenātim |\tke-na-tim |\tša kēnātim|\n|Month 4|\tIV |\tMahur-ilī |\tMa-hu-ur-DINGIR; ma-ḫu-ur-ì-lí |\t|\n|Month 5|\tV |\tAbšarrani |\táb-ša-ra-ni; áb ša-ra-ni; áb-ša-ra-nu |\tab šarrāni; abšarrani |\n|Month 6|\tVI |\tHubur |\tHu-bu-ur\t| |\n|Month 7|\tVII |\tṢip'um |\tṣí-ip-im |\tṣipum|\n|Month 8|\tVIII |\tQarrātum |\tqá-ra-a-tí; qá-ra-a-tim |\t|\n|Month 9|\tIX |\tKanwarta |\tkán-bar-ta; Kà-an-ma-ar-ta |\tKanmarta |\n|Month 10|\tX |\tTe’inātum |\tté-i-na-tim |\t|\n|Month 11|\tXI |\tKuzallum |\tku-zal-li; ku-zal-lu |\t|\n|Month 12|\tXII |\tAllanātum |\ta-lá-na-tum; a-lá-na-tim |\t|\n\n\n**Optional changes in transliterations:**\n- `Ḫ` → `H`\n- `ḫ` → `h`\n- `KÙ.B.` —> `KÙ.BABBAR`\n\n**Change unicode subscript numbers to normal integers in transliterations:**\n- `₀` → `0`\n- `₁` → `1` \n- `₂` → `2`\n- `₃` → `3`\n- `₄` → `4`\n- `₅` → `5`\n- `₆` → `6`\n- `₇` → `7`\n- `₈` → `8`\n- `₉` → `9`\n\n**Decimals to Fractions for transliterations:**\n- `0.5`  —> `½`\n- `0.25` —> `¼`\n- `0.3333` —> `⅓`\n- `0.8333` —> `⅚`\n- `0.625` —> `⅝`\n- `0.6666` —> `⅔`\n- `0.75` —>  `¾`\n- `0.1666` —> `⅙`\n\n**Examples of the Outcomes for Test**\n\n|Transliteration|Translation|\n|-----------|---------------|\n| `1 e-ma-ar-šu <gap>`| `1 donkey of his <gap>`|\n| `<gap> a-na Ú-<gap> a-dí-in`| `<gap> I gave it to U-<gap>.`|\n| `<gap> ⅓ ma-na a-na En-na-nim DUMU Am-ri-a áš-qúl`| `I paid <gap> ⅓ mina (silver) to Ennānum, son of Amriya.`|\n\n**Examples of the Optimal Outcomes for Train**\n\n|Transliteration|Translation|\n|-----------|---------------|\n|`<gap> ma-na KÙ.BABBAR ṣa-ru-pá-am <gap> GÍN KÙ.GI pá-ša-lam` | `<gap> minas of refined silver, <gap> shekels of pašallum gold` |\n|`<gap> GÚ SÍG.HI.A <gap> 5 maš-ku <gap> 22 na-ru-qá-tum 4 ANŠE ṣa-la-mu`|`<gap> talents of wool, <gap> 5 hides, <gap> 22 sacks, 4 black donkeys`|\n|`⅓ ma-na 2 ½ GÍN KÙ.BABBAR 20 NINDA i-ṣé-er tù-wa-ra-a-ah-šu a-lá-hu-um i-šu` | `Tuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.`|",
    "3414072": "@deeppast Thanks for the update. Some clarification to be 101% sure:\n\n**About double quotes:**\n\n>Do not remove from translations (as these are in the test too):\n\n>quotation marks \" \"\n\nIn the previous update:\n\n>quotations \" “= removed\n\nThe final rule is keep all quotation marks, correct? And for curly double quotes, we have to convert them to plain double quotes or not?\n\n**About  parentheses:**\n\nIn the previous update:\n>parentheses ( )= removed\n\nNow, we need to keep parenthesis as for the examples outcomes you've provided, right?\n\nFor instance: `0faa50f7-b86c-466c-a8ab-3a6f48fcb00a`\n\n\"at 52.5 grains (of silver) each, 0.5 shekel\" we must keep the parenthesis around `of silver`, right?\n\nWe must not do: \"at 52.5 grains of silver each, 0.5 shekel\"\n\n**About replacements:**\n>-tax —> šadduātum tax\n\n>-gold —> pašallum gold\n\n>textiles —> kutānum textiles (are you sure it's not -textiles —> kutānum textiles )?\n\nAnd it's `textiles` and `-tax` and `-gold` surrounded by spaces? \nIf not then in `198e428d-f51b-40d1-96d8-aee4bfa60d8d`:\n\n\"of Ennam-Suen, 8 textiles as **import-tax,** 10 textiles are pre-emption\" \n\nwould become: \n\n\"of Ennam-Suen, 8 kutānum textiles as **importšadduātum tax**, 10 kutānum textiles are pre-emption\". \n\nand it 229be03b-c772-4a6c-8792-c3f80948a97d:\n\n\"received 28 kutānu-textiles and 1 black donkey\"\n\nwould become: \n\n\"received 28 **kutānu-kutānum textiles** and 1 black donkey\"\n\nWhich would look weird, please confirm.\n\n**About `<gap>`:**\n\nIn previous update:\n>added space around gap in translations (not transliterations)\n\nIs it still true? Do we need to add space around `<gap>` except if glued with a dash?\n",
    "3416050": "Thanks for the comprehensive update and all the clarifications in the comments. After applying the v3 changes, a few things that helped on my end:\n\nRecovering truncated transliterations: Confirmed that ~10% are affected. Matching oare_id against published_texts.csv and cross-referencing with the AKT PDFs recovered most of them. Be careful, though,  some entries aren't just truncated; the translation doesn't match the transliteration at all (as MPWARE pointed out for AKT 8, 55), so it's worth doing a length-ratio sanity check before blindly patching.\n\nThe -textiles / -gold / -tax replacements: Based on the host's clarification, these only apply when preceded by a space (i.e., -textiles to kutānum textiles), not in the middle of a word, like import-tax or kutānu-textiles. A naive *str.replace* will break things, regex with word boundaries or explicit space matching is safer here.\n\nStraight quotes: Curly quotes to straight quotes for both \" and '. Small thing, but easy to miss if your text editor or PDF extraction reintroduces curly ones.\n\nStill working through the fraction conversions and subscript normalization. Good luck to everyone in the final stretch!",
    "3414172": "I think you should update these data from the beginning.Some samples from train.csv were aligned by hand. To revise these one more is really a challenging work.",
    "3414169": "@deeppast  hihi, \n\nRegarding the unit conversions that you mentioned below, I suspect that the right-hand-side of the first two rows should be interchanged\n\n```plaintext\n1 / 12 (shekel) —> ⅔ shekel 15 grains\n5 / 12 shekel —> 15 grains\n5 11 / 12 shekels —> 6 shekels less 15 grains\n7 / 12 shekel —> ½ shekel 15 grains\n```\n\nIs it that `1/12 shekel == 15 grains` ?\n\nif so, then the conversions should instead be \n\n```plaintext\n1 / 12 (shekel) —> 15 grains\n5 / 12 shekel —> ⅔ shekel 15 grains\n5 11 / 12 shekels —> 6 shekels less 15 grains\n7 / 12 shekel —> ½ shekel 15 grains\n```\n\nis it?",
    "3414139": "I am a bit confused about the table that follows 'Results of these changes: no more `<big_gap>`'. That table still says `<gap> <gap>: 30`, `<gap> <big_gap>: 6`, etc. Does this mean that there's till 30 occurrences of `<gap> <gap>`, 6 occurrences of `<gap> <big_gap>`, etc. in the test set?\n\nI thought I had understood that all of the `<big_gap>` had been replaced by `<gap>` and neighboring gaps had been merged together, but I'm less certain I understand right now.\n\nEdit: it seems that some time since this post `Results of these changes: no more <big_gap>` was changed to `Results of these changes: no more <big_gap> and no more duplicates for <gap> Before the change took place, these types of gaps were found in the test and training data. This recent change reduced all these duplicates to a single <gap>.` so it's now clearer. Thank you for the clarification!",
    "3418958": "Sorry if I am being dumb here\n\n```\nThe rest is up to you. Here are some recommendations, but there's no guarantee that these will improve your score on the leader board:\n```\n\nwhy is it “no guarantee”? IMO these are deterministic postprocessing, so if the test set conforms to this, the score will 100% improve. Does this mean only some texts were changed in this manner while the others are left as it is?",
    "3415113": "It looks like the “Change Roman numerals to integer numbers for months” table is misaligned/broken: Month 2 is labeled as Roman III, and III appears twice. \n\nAccording to Kouwenberg (Introduction to Old Assyrian, p. 182), the Old Assyrian calendar starts with Month 1 = I (Bēlat ekallem/Bēltekallem), followed by Month 2 = II (ša sarrātim) and Month 3 = III (ša kēnātim).\nCould you add Month 1 and shift/correct the Roman numerals accordingly?",
    "3414479": "The new update seems to have truncated the longer transliteration strings. The previous training data had the following long training entries:\n\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>592</td>\n      <td>746</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>338</td>\n      <td>588</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>343</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>408</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>357</td>\n      <td>626</td>\n    </tr>\n    <tr>\n      <td>1260</td>\n      <td>c97bb594-a5a1-4674-9496-48496e91c2ee</td>\n      <td>29</td>\n      <td>602</td>\n    </tr>\n    <tr>\n      <td>1378</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>379</td>\n      <td>603</td>\n    </tr>\n  </tbody>\n</table>\n\n<br>\n**The new data has truncated these to be:**\n\n<table>\n  <thead>\n    <tr>\n      <th>index</th>\n      <th>oare_id</th>\n      <th>transliteration_length</th>\n      <th>translation_length</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <td>177</td>\n      <td>1c428f5b-15b5-463a-8e27-f9b2f0d858fc</td>\n      <td>138</td>\n      <td>739</td>\n    </tr>\n    <tr>\n      <td>647</td>\n      <td>64f3382c-cb81-41af-99c9-36854585b747</td>\n      <td>107</td>\n      <td>587</td>\n    </tr>\n    <tr>\n      <td>927</td>\n      <td>937e71fb-57c8-46c0-afbc-48f9429886e4</td>\n      <td>114</td>\n      <td>537</td>\n    </tr>\n    <tr>\n      <td>968</td>\n      <td>991eef40-f139-4206-90ed-7fb45648b197</td>\n      <td>108</td>\n      <td>593</td>\n    </tr>\n    <tr>\n      <td>1090</td>\n      <td>adb0573b-20fb-469d-8343-0ace8e2489e0</td>\n      <td>107</td>\n      <td>625</td>\n    </tr>\n    <tr>\n      <td>1377</td>\n      <td>dff850c8-ccd4-44a9-9994-2834ca832a6d</td>\n      <td>120</td>\n      <td>585</td>\n    </tr>\n  </tbody>\n</table>\n\n\n<br>\nI'm not sure if this applies to the other rows too, but it certainly seems to have lost a lot of data.",
    "3416596": "question about (n broken lines), do the lines between \"Nimar-Istar.\" and \"When he returns from Mamma he will bring\nit to me\" counts as n broken lines?\n\nIf I understand correctly, the translation should be:\n\nI furthermore gave 1 mina of good, native copper for an allu’āru-container of sweet wine from Mamma to Puzur-Amurrum son of Nimar-Ištar. \\<gap> When he returns from Mamma he will bring it to me. (This was) apart from the 4 allu’āru-containers, the proceeds from the silver and washed copper that they have received (previously). Witnessed by Šu-Bēlum son of Kuzizia.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21691005%2F5a2fcafea533880783e26906cb8cdffd%2FScreenshot%202026-03-03%20174635.png?generation=1772534814707795&alt=media)",
    "3414925": "Regarding the terms \"pašallum gold\", \"šadduātum tax\", \"kutānum textiles\", those appear in different forms in the texts. Are those the forms that are expected from the test data (e.g. \"kutanum-textile\", \"kutanu textile\" -> \"kutānum textiles\")?",
    "3414179": "just for the record, **this update does not include all the changes made in the previous update**  \nFor example unicode subscript numbers are now present in the train transliteration:   \t\n`0064939c-59b9-4448-a63d-34612af0a1b5 -> 1 TÚG ša qá-tim i-tur₄-DINGIR il₅-qé ...`",
    "3413998": "Thanks, these are some new changes. Do we expect another test set update?",
    "3415275": "I'm trying to identify the mapping between oare_id in train.csv and the AKT PDF in order to complete broken sections, I've found 1561 matches:\n- AKT 5 = 77 oare_id \n- AKT 6a = 304 oare_id\n- AKT 6b = 222 oare_id\n- AKT 6c = 201 oare_id\n- AKT 6d = 139 oare_id\n- AKT 6e = 255 oare_id\n- AKT 8 = 363 oare_id (e.g. `0123a9b9-e81e-4d7a-a79b-10e7c0aacbb9` Kt 91/k 471, page 213)\n\nSomeone with similar results? @jackvd You was saying around 400 items was broken?",
    "3414364": "Can't describe an amount of confusion when comparing a new train dropped last week and a new train from yesterday and now I'm trying do understand how to merge those two\n\nupd: sorry, my humble verdict is that these updates last week made things worse...",
    "3414173": "```\nRemove from translations:\n\n    fem.\n    sing.\n    pl.\n    plural\n    (?)\n    any stray marks you find (e.g., .., ?, x, xx, << >>, < > except around <gap> of course)\n```\nYou recommend removing these words but would these words be included in the test set?",
    "3414058": "The updated data is now live on the site. I will commence the rescore of existing submissions shortly.\n\nUPDATE 02/26/2026: The rescore is now complete. Please let us know if you have any questions or concerns.",
    "3424553": "Cool ,its intresting ",
    "3421313": "Sorry, I still don't understand what a \"meaningful\" question mark and exclamation point are. If they're enclosed in parentheses with words like: (and?), should I change that to -> (and) or leave them as is?",
    "3415409": "1 / 12 (shekel) —> 15 grains\n\nwhat should we do with 1/4, 1/6, and 1/8",
    "3415296": "@deeppast  hi, just to be 100% clarified, is ḫ gonna be a valid character in the test set? \n\nAsking this because in the \"Examples of the Optimal Outcomes for Train\" you mentioned, the proper noun \"Tuwar-a**ḫ**šu\" is displayed as \"Tuwar-a**h**šu\" \n\nthanks!",
    "3414239": "Hi! \nCan you clarify these examples in translation:\n\n```python\nI paid <gap> ⅓ mina (silver) to Ennānum, son of Amriya.\nTuwar-ahšu owes ⅓ mina 2 ½ shekels of silver (and) 20 loaves of bread to Ali-ahum.\n```\n\nWhy are we using curved parenthesis? I looked at train.csv on datasets page and there are no examples of \"(silver)\". Or examples of parenthesis usage.",
    "3414137": "Thanks for the update! \n\nMay I ask about a specific case: 33773ec0-e74f-41bf-b985-f3e35b0f26c9\n\n```plaintext\nFrom what Amur-Aššur (owes?)\n```\nwill such case of `(some word?)`  appear in the test set? ",
    "3413999": "Thanks @deeppast for your ongoing efforts. What does this exactly mean?\n> PN —> < gap >\n\nDoes this mean that every personal name to be found must be replaced by a gap tag? Or just the literal `PN`?\n\nAlso, \n> Shortening of long floats to four places after the decimal\n\nI am a bit confused by this since later you suggest converting them to unicode fractions. Does the hidden test set (both translations and transliterations) contain floating point numbers or only unicode fractions?\n\nThanks in advance ",
    "3414094": ""
  }
}