{
  "id": 681431,
  "title": "train.csv translation don't align with pdf version",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/681431",
  "author_name": "",
  "post_date": "2026-03-14T22:42:04.180209700Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi all, get question regarding translation's utf encoding. </p>\n<p>For record oare_id=<strong>6cc9086f-59c5-4b5a-aedc-ba77e60aed57</strong>,  the provided translation is </p>\n<pre><code>Seal of Kanesh. To Kuliya, the-payer, our messenger, to every single, wherever Sukkaliya, son of Aššuriš-tikal is staying, small and big.\n</code></pre>\n<p>where the translation provided in \"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf?select=AKT+5+2008.pdf\" is </p>\n<pre><code>Seal of kārum Kanesh. To Kuliya, the datu-payer, our messenger, to every single kärum, wherever ukkaliya, son of Aššuriš-tikal is staying, small and big.\n</code></pre>\n<p>can someone from host help me figured out which is the expected version for training (align with test csv's expectation)?</p>",
  "messages": [
    {
      "id": "3421194",
      "postDate": "03/14/2026 22:42:04",
      "content": "<p>Hi all, get question regarding translation's utf encoding. </p>\n<p>For record oare_id=<strong>6cc9086f-59c5-4b5a-aedc-ba77e60aed57</strong>,  the provided translation is </p>\n<pre><code>Seal of Kanesh. To Kuliya, the-payer, our messenger, to every single, wherever Sukkaliya, son of Aššuriš-tikal is staying, small and big.\n</code></pre>\n<p>where the translation provided in \"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf?select=AKT+5+2008.pdf\" is </p>\n<pre><code>Seal of kārum Kanesh. To Kuliya, the datu-payer, our messenger, to every single kärum, wherever ukkaliya, son of Aššuriš-tikal is staying, small and big.\n</code></pre>\n<p>can someone from host help me figured out which is the expected version for training (align with test csv's expectation)?</p>",
      "rawMarkdown": "Hi all, get question regarding translation's utf encoding. \n\nFor record oare_id=**6cc9086f-59c5-4b5a-aedc-ba77e60aed57**,  the provided translation is \n```\nSeal of Kanesh. To Kuliya, the-payer, our messenger, to every single, wherever Sukkaliya, son of Aššuriš-tikal is staying, small and big.\n```\nwhere the translation provided in \"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf?select=AKT+5+2008.pdf\" is \n```\nSeal of kārum Kanesh. To Kuliya, the datu-payer, our messenger, to every single kärum, wherever ukkaliya, son of Aššuriš-tikal is staying, small and big.\n```\n\ncan someone from host help me figured out which is the expected version for training (align with test csv's expectation)?",
      "votes": null
    },
    {
      "id": "3421257",
      "postDate": "03/15/2026 03:30:25",
      "content": "<p>Yes, and what's not ideal is that the metrics used reward exact ngram overlap rather than semantic closeness. Factoring in a fuzzier similarity metric like the cosine distance between the reference translation and what's submitted could have helped mitigate some issues, but it's not what was chosen. The data changed a few times along the way as well, and it is still not entirely clear whether the training set provided matches the style of the test set or whether the data cleaning instructions given help maximize ngram overlap or are ideals aimed at improving translation quality in a general sense rather than matching the style of the actual test set.</p>",
      "rawMarkdown": "Yes, and what's not ideal is that the metrics used reward exact ngram overlap rather than semantic closeness. Factoring in a fuzzier similarity metric like the cosine distance between the reference translation and what's submitted could have helped mitigate some issues, but it's not what was chosen. The data changed a few times along the way as well, and it is still not entirely clear whether the training set provided matches the style of the test set or whether the data cleaning instructions given help maximize ngram overlap or are ideals aimed at improving translation quality in a general sense rather than matching the style of the actual test set.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3421257,
      "author_name": "epiktistes",
      "author_url": "",
      "post_date": "03/15/2026 03:30:25",
      "content": "<p>Yes, and what's not ideal is that the metrics used reward exact ngram overlap rather than semantic closeness. Factoring in a fuzzier similarity metric like the cosine distance between the reference translation and what's submitted could have helped mitigate some issues, but it's not what was chosen. The data changed a few times along the way as well, and it is still not entirely clear whether the training set provided matches the style of the test set or whether the data cleaning instructions given help maximize ngram overlap or are ideals aimed at improving translation quality in a general sense rather than matching the style of the actual test set.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3421194": "Hi all, get question regarding translation's utf encoding. \n\nFor record oare_id=**6cc9086f-59c5-4b5a-aedc-ba77e60aed57**,  the provided translation is \n```\nSeal of Kanesh. To Kuliya, the-payer, our messenger, to every single, wherever Sukkaliya, son of Aššuriš-tikal is staying, small and big.\n```\nwhere the translation provided in \"https://www.kaggle.com/datasets/deeppast/old-assyrian-kltepe-tablets-in-pdf?select=AKT+5+2008.pdf\" is \n```\nSeal of kārum Kanesh. To Kuliya, the datu-payer, our messenger, to every single kärum, wherever ukkaliya, son of Aššuriš-tikal is staying, small and big.\n```\n\ncan someone from host help me figured out which is the expected version for training (align with test csv's expectation)?",
    "3421257": "Yes, and what's not ideal is that the metrics used reward exact ngram overlap rather than semantic closeness. Factoring in a fuzzier similarity metric like the cosine distance between the reference translation and what's submitted could have helped mitigate some issues, but it's not what was chosen. The data changed a few times along the way as well, and it is still not entirely clear whether the training set provided matches the style of the test set or whether the data cleaning instructions given help maximize ngram overlap or are ideals aimed at improving translation quality in a general sense rather than matching the style of the actual test set."
  },
  "source": "meta"
}