{
  "id": 669151,
  "title": "Training from scratch vs finetuning existing model",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/669151",
  "author_name": "",
  "post_date": "2026-01-20T20:48:41.551391900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Do u guys think it's better to train translation model from scratch or finetune large llm model? Not sure how much benefit u can extract from training pretrained model, I might assume that translating akkadian isn't specialty of llms?</p>",
  "messages": [
    {
      "id": "3394325",
      "postDate": "01/20/2026 20:48:41",
      "content": "<p>Do u guys think it's better to train translation model from scratch or finetune large llm model? Not sure how much benefit u can extract from training pretrained model, I might assume that translating akkadian isn't specialty of llms?</p>",
      "rawMarkdown": "Do u guys think it's better to train translation model from scratch or finetune large llm model? Not sure how much benefit u can extract from training pretrained model, I might assume that translating akkadian isn't specialty of llms?",
      "votes": null
    },
    {
      "id": "3394429",
      "postDate": "01/21/2026 02:41:31",
      "content": "<p>The mainstream approach by far is to finetune byt5, a llm model. I'm not sure what you mean by training from scratch but it would be extremely hard to collect enough Akkadian text for that. (maybe you mean finetuning a shared competition model?)</p>",
      "rawMarkdown": "The mainstream approach by far is to finetune byt5, a llm model. I'm not sure what you mean by training from scratch but it would be extremely hard to collect enough Akkadian text for that. (maybe you mean finetuning a shared competition model?)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3394429,
      "author_name": "joey111mygo",
      "author_url": "",
      "post_date": "01/21/2026 02:41:31",
      "content": "<p>The mainstream approach by far is to finetune byt5, a llm model. I'm not sure what you mean by training from scratch but it would be extremely hard to collect enough Akkadian text for that. (maybe you mean finetuning a shared competition model?)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3394325": "Do u guys think it's better to train translation model from scratch or finetune large llm model? Not sure how much benefit u can extract from training pretrained model, I might assume that translating akkadian isn't specialty of llms?",
    "3394429": "The mainstream approach by far is to finetune byt5, a llm model. I'm not sure what you mean by training from scratch but it would be extremely hard to collect enough Akkadian text for that. (maybe you mean finetuning a shared competition model?)"
  },
  "source": "meta"
}