{
  "id": 669544,
  "title": "Should we use pre-trained models or train them ourselves?",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/669544",
  "author_name": "",
  "post_date": "2026-01-22T23:01:38.113927800Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I am aware that the current strategy by most people is to use pre-trained byT5 models, but which is better: using those pre-trained models, or training byT5 on Akkadian data yourself? \nI am aware that you can achieve high scores by using models trained by other people, but there are people who have gotten extremely high scores, and I’m just wondering, which are they doing?</p>",
  "messages": [
    {
      "id": "3395409",
      "postDate": "01/22/2026 23:01:38",
      "content": "<p>I am aware that the current strategy by most people is to use pre-trained byT5 models, but which is better: using those pre-trained models, or training byT5 on Akkadian data yourself? \nI am aware that you can achieve high scores by using models trained by other people, but there are people who have gotten extremely high scores, and I’m just wondering, which are they doing?</p>",
      "rawMarkdown": "I am aware that the current strategy by most people is to use pre-trained byT5 models, but which is better: using those pre-trained models, or training byT5 on Akkadian data yourself? \nI am aware that you can achieve high scores by using models trained by other people, but there are people who have gotten extremely high scores, and I’m just wondering, which are they doing?",
      "votes": null
    },
    {
      "id": "3395454",
      "postDate": "01/23/2026 02:31:02",
      "content": "<p>One issue with currently-shared models is that the preprocessing approaches differ. They handle the special characters, gaps, etc. in diverse ways, and most of them use the raw train.csv instead of an aligned dataset. Personally I would recommend training models yourself, as this is a good way of learning and at least you’re not at the mercy of whether someone decides to share their model😂.</p>",
      "rawMarkdown": "One issue with currently-shared models is that the preprocessing approaches differ. They handle the special characters, gaps, etc. in diverse ways, and most of them use the raw train.csv instead of an aligned dataset. Personally I would recommend training models yourself, as this is a good way of learning and at least you’re not at the mercy of whether someone decides to share their model😂.",
      "votes": null
    },
    {
      "id": "3395677",
      "postDate": "01/23/2026 13:06:54",
      "content": "<p>Yeah, Thats what I’ve been doing. But I wasn’t sure if there was a public model that was performing really well. Right now I’m fine-tuning byT5. Thanks for your help! </p>",
      "rawMarkdown": "Yeah, Thats what I’ve been doing. But I wasn’t sure if there was a public model that was performing really well. Right now I’m fine-tuning byT5. Thanks for your help!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3395454,
      "author_name": "joey111mygo",
      "author_url": "",
      "post_date": "01/23/2026 02:31:02",
      "content": "<p>One issue with currently-shared models is that the preprocessing approaches differ. They handle the special characters, gaps, etc. in diverse ways, and most of them use the raw train.csv instead of an aligned dataset. Personally I would recommend training models yourself, as this is a good way of learning and at least you’re not at the mercy of whether someone decides to share their model😂.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3395677,
          "author_name": "leowithrow",
          "author_url": "",
          "post_date": "01/23/2026 13:06:54",
          "content": "<p>Yeah, Thats what I’ve been doing. But I wasn’t sure if there was a public model that was performing really well. Right now I’m fine-tuning byT5. Thanks for your help! </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3395409": "I am aware that the current strategy by most people is to use pre-trained byT5 models, but which is better: using those pre-trained models, or training byT5 on Akkadian data yourself? \nI am aware that you can achieve high scores by using models trained by other people, but there are people who have gotten extremely high scores, and I’m just wondering, which are they doing?",
    "3395454": "One issue with currently-shared models is that the preprocessing approaches differ. They handle the special characters, gaps, etc. in diverse ways, and most of them use the raw train.csv instead of an aligned dataset. Personally I would recommend training models yourself, as this is a good way of learning and at least you’re not at the mercy of whether someone decides to share their model😂.",
    "3395677": "Yeah, Thats what I’ve been doing. But I wasn’t sure if there was a public model that was performing really well. Right now I’m fine-tuning byT5. Thanks for your help!"
  },
  "source": "meta"
}