{
  "id": 667878,
  "title": "Write-up / extra data",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/667878",
  "author_name": "",
  "post_date": "2026-01-14T12:53:21.871438800Z",
  "votes": 7,
  "comment_count": 4,
  "views": 0,
  "content": "<p>My thoughts and what worked for me. I am not an expert and I am sure this is known to most people. \nMore interesting, my finetune model is attached.</p>\n<p>I still have the LLM extracted data, I might try to improve its quality in the future, but time…</p>\n<p><a href=\"https://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data\" target=\"_blank\">https://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data</a></p>",
  "messages": [
    {
      "id": "3391125",
      "postDate": "01/14/2026 12:53:21",
      "content": "<p>My thoughts and what worked for me. I am not an expert and I am sure this is known to most people. \nMore interesting, my finetune model is attached.</p>\n<p>I still have the LLM extracted data, I might try to improve its quality in the future, but time…</p>\n<p><a href=\"https://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data\" target=\"_blank\">https://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data</a></p>",
      "rawMarkdown": "My thoughts and what worked for me. I am not an expert and I am sure this is known to most people. \nMore interesting, my finetune model is attached.\n\nI still have the LLM extracted data, I might try to improve its quality in the future, but time...\n\nhttps://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data",
      "votes": null
    },
    {
      "id": "3391457",
      "postDate": "01/15/2026 02:30:25",
      "content": "<p>That was amazing work. Could you add information about the model's LB scores, for example, a table comparing the results between NLLB and Byt5 models trained on the original data and your supplementary data?</p>",
      "rawMarkdown": "That was amazing work. Could you add information about the model's LB scores, for example, a table comparing the results between NLLB and Byt5 models trained on the original data and your supplementary data?",
      "votes": null
    },
    {
      "id": "3391576",
      "postDate": "01/15/2026 08:06:50",
      "content": "<p>Thanks. I do not have that data, but I was getting results around 22-24 with NLLB. I saw a general improvement when going to Byt5, it was based on a post by someone here, and I did not redo all the training with NLLB.  I think, my main issue at the moment is data driven rather than model dependent so I did not spend to much time evaluating NLLB vs Byt5.</p>",
      "rawMarkdown": "Thanks. I do not have that data, but I was getting results around 22-24 with NLLB. I saw a general improvement when going to Byt5, it was based on a post by someone here, and I did not redo all the training with NLLB.  I think, my main issue at the moment is data driven rather than model dependent so I did not spend to much time evaluating NLLB vs Byt5.",
      "votes": null
    },
    {
      "id": "3391587",
      "postDate": "01/15/2026 08:31:47",
      "content": "<p>Hello, could you please share your processed data, or share akk_pairs_final.csv, akkademia.csv, or acc_akkadian_english.csv?</p>",
      "rawMarkdown": "Hello, could you please share your processed data, or share akk_pairs_final.csv, akkademia.csv, or acc_akkadian_english.csv?",
      "votes": null
    },
    {
      "id": "3392100",
      "postDate": "01/16/2026 09:34:45",
      "content": "<p>I will do if I decide to drop further work on this, but I might still try to compete. You can find part of the data here <a href=\"https://github.com/gaigutherz/Akkademia/tree/master/NMT_input\" target=\"_blank\">https://github.com/gaigutherz/Akkademia/tree/master/NMT_input</a>\nYou want the .tr and .en files.</p>",
      "rawMarkdown": "I will do if I decide to drop further work on this, but I might still try to compete. You can find part of the data here https://github.com/gaigutherz/Akkademia/tree/master/NMT_input\nYou want the .tr and .en files.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3391457,
      "author_name": "phucthaiv02",
      "author_url": "",
      "post_date": "01/15/2026 02:30:25",
      "content": "<p>That was amazing work. Could you add information about the model's LB scores, for example, a table comparing the results between NLLB and Byt5 models trained on the original data and your supplementary data?</p>",
      "votes": null,
      "replies": [
        {
          "id": 3391576,
          "author_name": "haakonandresen",
          "author_url": "",
          "post_date": "01/15/2026 08:06:50",
          "content": "<p>Thanks. I do not have that data, but I was getting results around 22-24 with NLLB. I saw a general improvement when going to Byt5, it was based on a post by someone here, and I did not redo all the training with NLLB.  I think, my main issue at the moment is data driven rather than model dependent so I did not spend to much time evaluating NLLB vs Byt5.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3391587,
              "author_name": "qianyuu",
              "author_url": "",
              "post_date": "01/15/2026 08:31:47",
              "content": "<p>Hello, could you please share your processed data, or share akk_pairs_final.csv, akkademia.csv, or acc_akkadian_english.csv?</p>",
              "votes": null,
              "replies": [
                {
                  "id": 3392100,
                  "author_name": "haakonandresen",
                  "author_url": "",
                  "post_date": "01/16/2026 09:34:45",
                  "content": "<p>I will do if I decide to drop further work on this, but I might still try to compete. You can find part of the data here <a href=\"https://github.com/gaigutherz/Akkademia/tree/master/NMT_input\" target=\"_blank\">https://github.com/gaigutherz/Akkademia/tree/master/NMT_input</a>\nYou want the .tr and .en files.</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3391125": "My thoughts and what worked for me. I am not an expert and I am sure this is known to most people. \nMore interesting, my finetune model is attached.\n\nI still have the LLM extracted data, I might try to improve its quality in the future, but time...\n\nhttps://kaggle.com/writeups/haakonandresen/akkadian-to-english-time-travel-to-bad-data",
    "3391457": "That was amazing work. Could you add information about the model's LB scores, for example, a table comparing the results between NLLB and Byt5 models trained on the original data and your supplementary data?",
    "3391576": "Thanks. I do not have that data, but I was getting results around 22-24 with NLLB. I saw a general improvement when going to Byt5, it was based on a post by someone here, and I did not redo all the training with NLLB.  I think, my main issue at the moment is data driven rather than model dependent so I did not spend to much time evaluating NLLB vs Byt5.",
    "3391587": "Hello, could you please share your processed data, or share akk_pairs_final.csv, akkademia.csv, or acc_akkadian_english.csv?",
    "3392100": "I will do if I decide to drop further work on this, but I might still try to compete. You can find part of the data here https://github.com/gaigutherz/Akkademia/tree/master/NMT_input\nYou want the .tr and .en files."
  },
  "source": "meta"
}