From the published texts there's a link to the AI translation, where the competition mentioned it to be \"poorly translated\", but I believe it might still have some use.
\nHere's the link to the notebook and processing to get those data:
\nhttps://www.kaggle.com/code/zhangyue199/dpc-ai-translation-dataset/output?scriptVersionId=290276456
\nI do have some confusion on the data, the AI translation page includes multiple sides of the tablet, e.g. obverse, reverse.
\nBut the Akkadian there does not exactly match with the transliteration in the publised_texts, for which I am not sure how to match the transliteration to the AI english.
Appreciate if anyone can shed some lights.
", "messages": [ { "id": 3386931, "postDate": "2026-01-06T05:06:26.500Z", "content": "From the published texts there's a link to the AI translation, where the competition mentioned it to be \"poorly translated\", but I believe it might still have some use.
\nHere's the link to the notebook and processing to get those data:
\nhttps://www.kaggle.com/code/zhangyue199/dpc-ai-translation-dataset/output?scriptVersionId=290276456
\nI do have some confusion on the data, the AI translation page includes multiple sides of the tablet, e.g. obverse, reverse.
\nBut the Akkadian there does not exactly match with the transliteration in the publised_texts, for which I am not sure how to match the transliteration to the AI english.
Appreciate if anyone can shed some lights.
", "rawMarkdown": "From the published texts there's a link to the AI translation, where the competition mentioned it to be \"poorly translated\", but I believe it might still have some use.\n\nHere's the link to the notebook and processing to get those data:\n\nhttps://www.kaggle.com/code/zhangyue199/dpc-ai-translation-dataset/output?scriptVersionId=290276456\n\nI do have some confusion on the data, the AI translation page includes multiple sides of the tablet, e.g. obverse, reverse.\n\nBut the Akkadian there does not exactly match with the `transliteration` in the `publised_texts`, for which I am not sure how to match the `transliteration` to the `AI english`.\n\nAppreciate if anyone can shed some lights.", "votes": 11 }, { "id": 3387129, "postDate": "2026-01-06T12:34:25.143Z", "content": "Just map them via the cdli_id. Although your current notebook doesn’t have a separate column for cdli_id, the cdli_id is included in url_org — you can extract it. Alternatively, you can rewrite the notebook to add a cdli_id column, which will allow matching with the content in published_texts.csv.\nAdditional Notes:
\naefa478d-c47d-42d2-a3b5-3f150603ccee\n55b115ac-6a6d-48e6-86df-ff449b4b346b\n99923b12-2acd-4eb4-9e1a-56b85c475a76\n3e87aad8-cd8c-4f8d-96e9-c68e062d4a1a\n4beda101-a827-47e2-b47a-799fd8f68c07\n8f46f357-9971-427f-ac9a-7de23c9ce3a5\nhi, even with the CDLI_ID, how would that help matching tranliteration to the AI translation?
", "rawMarkdown": "hi, even with the CDLI_ID, how would that help matching tranliteration to the AI translation?", "replies": [ { "id": 3387474, "postDate": "2026-01-07T05:06:44.123Z", "content": "The only viable method might be to use \nobverse_akkadian obverse_english\nreverse_akkadian reverse_english
and then standardize the Akkadian text back to the format specified for this competition.
", "rawMarkdown": "The only viable method might be to use \n`obverse_akkadian obverse_english\nreverse_akkadian reverse_english`\n\nand then standardize the Akkadian text back to the format specified for this competition." } ] } ] }, { "id": 3386994, "postDate": "2026-01-06T07:22:07.413Z", "content": "I think the transliteration_org column is similar to that in the train.csv. And the English translation is in obverse_english and reserve_english columns. Unfortunately, these two columns contain only part of the translation, and you may deduce the corresponding transliteration extract from the two akkadian columns (these are another form of expressing akkadian I believe, possibly CDLI).
\nFor instance, in the first row,
\n=>\nSeal of Ashur-malik, son of Ina, to Ashur-nada, son of … and Enum-assur, son of …, Aputum, in accordance with the wording of the tablet …, who .
\n=>\n. . . a scepter .
\nIn the second row, only the reserve_english part is valid:
\n=>\nFrom the mushtu-tree of Shubelum and to the 8 mushtu-trees he shall weigh out. That is what at that time
\n(this translation is terrible. It probably should be: From the week of Šū-bēlum, son of Anali, he shall pay in 8 weeks.)
", "rawMarkdown": "I think the transliteration_org column is similar to that in the train.csv. And the English translation is in obverse_english and reserve_english columns. Unfortunately, these two columns contain only part of the translation, and you may deduce the corresponding transliteration extract from the two akkadian columns (these are another form of expressing akkadian I believe, possibly CDLI).\n\nFor instance, in the first row,\n\n- KIŠIB a-šùr-ma-lik DUMU i-na-a a-na a-šùr-na-da DUMU … ù en-um-a-šùr DUMU … a-pu-tum a-na a-wa-at ṭup-pì-im iḫ-da … ša ú-…\n\n=>\nSeal of Ashur-malik, son of Ina, to Ashur-nada, son of ... and Enum-assur, son of ..., Aputum, in accordance with the wording of the tablet ..., who .\n\n- ḫa-sí-sú …\n\n=>\n. . . a scepter .\n\nIn the second row, only the reserve_english part is valid:\n\n- iš-tù ḫa-mu-uš-tim ša šu-be-lim DUMU a-na-lí a-na 8 ḫa-am-ša-tim i-ša-qal\n\n=>\nFrom the mushtu-tree of Shubelum and to the 8 mushtu-trees he shall weigh out. That is what at that time\n\n(this translation is terrible. It probably should be: From the week of Šū-bēlum, son of Anali, he shall pay in 8 weeks.)", "votes": 2, "replies": [ { "id": 3387117, "postDate": "2026-01-06T12:07:09.733Z", "content": "yeah, but the Akkadian from the link seems to be a different Akkadian from the training example, I think using these are training would deteriorite the model.
\nMy guess is it would work only if we can find the AI translations that matches the transliteration_org in the dataset.
Yes, it's not the form this competition expects. As I mentioned, the two columns could only serve to locate their counterparts in the transliteration_org column.
\nThat said, the Akkadian columns are far from good splits of the transliteration_org documents. If time permits, you may build the aligned database through more advanced LLMs which could both split the sentences and better translate them.
", "rawMarkdown": "Yes, it's not the form this competition expects. As I mentioned, the two columns could only serve to locate their counterparts in the transliteration_org column. \n\nThat said, the Akkadian columns are far from good splits of the transliteration_org documents. If time permits, you may build the aligned database through more advanced LLMs which could both split the sentences and better translate them.", "votes": 1 } ] } ] }, { "id": 3387127, "postDate": "2026-01-06T12:33:04.193Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3387129, "author_name": "耶✌", "author_url": "", "post_date": "2026-01-06T12:34:25.143000", "content": "Just map them via the cdli_id. Although your current notebook doesn’t have a separate column for cdli_id, the cdli_id is included in url_org — you can extract it. Alternatively, you can rewrite the notebook to add a cdli_id column, which will allow matching with the content in published_texts.csv.\nAdditional Notes:
\naefa478d-c47d-42d2-a3b5-3f150603ccee\n55b115ac-6a6d-48e6-86df-ff449b4b346b\n99923b12-2acd-4eb4-9e1a-56b85c475a76\n3e87aad8-cd8c-4f8d-96e9-c68e062d4a1a\n4beda101-a827-47e2-b47a-799fd8f68c07\n8f46f357-9971-427f-ac9a-7de23c9ce3a5\nhi, even with the CDLI_ID, how would that help matching tranliteration to the AI translation?
", "votes": 0, "replies": [ { "id": 3387474, "author_name": "耶✌", "author_url": "", "post_date": "2026-01-07T05:06:44.123000", "content": "The only viable method might be to use \nobverse_akkadian obverse_english\nreverse_akkadian reverse_english
and then standardize the Akkadian text back to the format specified for this competition.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3386994, "author_name": "Tomorin", "author_url": "", "post_date": "2026-01-06T07:22:07.413000", "content": "I think the transliteration_org column is similar to that in the train.csv. And the English translation is in obverse_english and reserve_english columns. Unfortunately, these two columns contain only part of the translation, and you may deduce the corresponding transliteration extract from the two akkadian columns (these are another form of expressing akkadian I believe, possibly CDLI).
\nFor instance, in the first row,
\n=>\nSeal of Ashur-malik, son of Ina, to Ashur-nada, son of … and Enum-assur, son of …, Aputum, in accordance with the wording of the tablet …, who .
\n=>\n. . . a scepter .
\nIn the second row, only the reserve_english part is valid:
\n=>\nFrom the mushtu-tree of Shubelum and to the 8 mushtu-trees he shall weigh out. That is what at that time
\n(this translation is terrible. It probably should be: From the week of Šū-bēlum, son of Anali, he shall pay in 8 weeks.)
", "votes": 2, "replies": [ { "id": 3387117, "author_name": "MJeremy", "author_url": "", "post_date": "2026-01-06T12:07:09.733000", "content": "yeah, but the Akkadian from the link seems to be a different Akkadian from the training example, I think using these are training would deteriorite the model.
\nMy guess is it would work only if we can find the AI translations that matches the transliteration_org in the dataset.
Yes, it's not the form this competition expects. As I mentioned, the two columns could only serve to locate their counterparts in the transliteration_org column.
\nThat said, the Akkadian columns are far from good splits of the transliteration_org documents. If time permits, you may build the aligned database through more advanced LLMs which could both split the sentences and better translate them.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3387127, "author_name": "", "author_url": "", "post_date": "2026-01-06T12:33:04.193000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3386931": "From the published texts there's a link to the AI translation, where the competition mentioned it to be \"poorly translated\", but I believe it might still have some use.\n\nHere's the link to the notebook and processing to get those data:\n\nhttps://www.kaggle.com/code/zhangyue199/dpc-ai-translation-dataset/output?scriptVersionId=290276456\n\nI do have some confusion on the data, the AI translation page includes multiple sides of the tablet, e.g. obverse, reverse.\n\nBut the Akkadian there does not exactly match with the `transliteration` in the `publised_texts`, for which I am not sure how to match the `transliteration` to the `AI english`.\n\nAppreciate if anyone can shed some lights.", "3387129": "Just map them via the **cdli_id**. Although your current notebook doesn’t have a separate column for cdli_id, the cdli_id is included in **url_org** — you can extract it. Alternatively, you can rewrite the notebook to add a cdli_id column, which will allow matching with the content in published_texts.csv.\nAdditional Notes:\n1. The Akkadian content and AI-translated content on the original website have **multiple sides**, but this notebook only scrapes the \"Obverse\" side. For reference, the following oare_ids are the only ones from **train.csv that appear on the website**:\n```python\naefa478d-c47d-42d2-a3b5-3f150603ccee\n55b115ac-6a6d-48e6-86df-ff449b4b346b\n99923b12-2acd-4eb4-9e1a-56b85c475a76\n3e87aad8-cd8c-4f8d-96e9-c68e062d4a1a\n4beda101-a827-47e2-b47a-799fd8f68c07\n8f46f357-9971-427f-ac9a-7de23c9ce3a5\n```\n2. The transliteration on this website uses the **CDLI notation**, which differs from that of the competition. For example, this website uses \"a2\" instead of \"á\", along with other variations in capitalization and determiners — normalization is required.", "3386994": "I think the transliteration_org column is similar to that in the train.csv. And the English translation is in obverse_english and reserve_english columns. Unfortunately, these two columns contain only part of the translation, and you may deduce the corresponding transliteration extract from the two akkadian columns (these are another form of expressing akkadian I believe, possibly CDLI).\n\nFor instance, in the first row,\n\n- KIŠIB a-šùr-ma-lik DUMU i-na-a a-na a-šùr-na-da DUMU … ù en-um-a-šùr DUMU … a-pu-tum a-na a-wa-at ṭup-pì-im iḫ-da … ša ú-…\n\n=>\nSeal of Ashur-malik, son of Ina, to Ashur-nada, son of ... and Enum-assur, son of ..., Aputum, in accordance with the wording of the tablet ..., who .\n\n- ḫa-sí-sú …\n\n=>\n. . . a scepter .\n\nIn the second row, only the reserve_english part is valid:\n\n- iš-tù ḫa-mu-uš-tim ša šu-be-lim DUMU a-na-lí a-na 8 ḫa-am-ša-tim i-ša-qal\n\n=>\nFrom the mushtu-tree of Shubelum and to the 8 mushtu-trees he shall weigh out. That is what at that time\n\n(this translation is terrible. It probably should be: From the week of Šū-bēlum, son of Anali, he shall pay in 8 weeks.)", "3387127": "" } }