{
  "id": 669055,
  "title": "Does this competition not require training models?",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/669055",
  "author_name": "",
  "post_date": "2026-01-20T11:43:25.450866900Z",
  "votes": -1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I looked at other people's notebooks, and none of them involved model training; they all generated directly.</p>",
  "messages": [
    {
      "id": "3394069",
      "postDate": "01/20/2026 11:43:25",
      "content": "<p>I looked at other people's notebooks, and none of them involved model training; they all generated directly.</p>",
      "rawMarkdown": "I looked at other people's notebooks, and none of them involved model training; they all generated directly.",
      "votes": null
    },
    {
      "id": "3394073",
      "postDate": "01/20/2026 11:57:26",
      "content": "<p>What you see are copy-pasted notebooks with ensembles of models (trained and released open-sourse by some contestants like 2-3 weeks ago) with +-0.001 weight tuning in hopes of achieving +5-7 positions on the LB or get some medals for notebooks~~~~</p>\n<p>If you are interested in train — <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-train\" target=\"_blank\">that</a> could be a good starter baseline.</p>",
      "rawMarkdown": "What you see are copy-pasted notebooks with ensembles of models (trained and released open-sourse by some contestants like 2-3 weeks ago) with +-0.001 weight tuning in hopes of achieving +5-7 positions on the LB or get some medals for notebooks~~~~\n\nIf you are interested in train — [that](https://www.kaggle.com/code/takamichitoda/dpc-starter-train) could be a good starter baseline.",
      "votes": null
    },
    {
      "id": "3394275",
      "postDate": "01/20/2026 18:58:43",
      "content": "<p>lol! What you said cracks me up. Yeah, most of the code right now is model ensembling. You can check out the code titled train. Currently, the ByT5 model is what most people are using, but you can give other models a try.</p>",
      "rawMarkdown": "lol! What you said cracks me up. Yeah, most of the code right now is model ensembling. You can check out the code titled train. Currently, the ByT5 model is what most people are using, but you can give other models a try.",
      "votes": null
    },
    {
      "id": "3394427",
      "postDate": "01/21/2026 02:34:34",
      "content": "<p>The models shared only serve as the baseline -- they could be used for further finetuning, ensemble, or just for testing a preprocess idea. However, to succeed in the competition you need more than rerun these inference notebooks.</p>",
      "rawMarkdown": "The models shared only serve as the baseline -- they could be used for further finetuning, ensemble, or just for testing a preprocess idea. However, to succeed in the competition you need more than rerun these inference notebooks.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3394073,
      "author_name": "samson8",
      "author_url": "",
      "post_date": "01/20/2026 11:57:26",
      "content": "<p>What you see are copy-pasted notebooks with ensembles of models (trained and released open-sourse by some contestants like 2-3 weeks ago) with +-0.001 weight tuning in hopes of achieving +5-7 positions on the LB or get some medals for notebooks~~~~</p>\n<p>If you are interested in train — <a href=\"https://www.kaggle.com/code/takamichitoda/dpc-starter-train\" target=\"_blank\">that</a> could be a good starter baseline.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3394275,
      "author_name": "qifeihhh666",
      "author_url": "",
      "post_date": "01/20/2026 18:58:43",
      "content": "<p>lol! What you said cracks me up. Yeah, most of the code right now is model ensembling. You can check out the code titled train. Currently, the ByT5 model is what most people are using, but you can give other models a try.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3394427,
      "author_name": "joey111mygo",
      "author_url": "",
      "post_date": "01/21/2026 02:34:34",
      "content": "<p>The models shared only serve as the baseline -- they could be used for further finetuning, ensemble, or just for testing a preprocess idea. However, to succeed in the competition you need more than rerun these inference notebooks.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3394069": "I looked at other people's notebooks, and none of them involved model training; they all generated directly.",
    "3394073": "What you see are copy-pasted notebooks with ensembles of models (trained and released open-sourse by some contestants like 2-3 weeks ago) with +-0.001 weight tuning in hopes of achieving +5-7 positions on the LB or get some medals for notebooks~~~~\n\nIf you are interested in train — [that](https://www.kaggle.com/code/takamichitoda/dpc-starter-train) could be a good starter baseline.",
    "3394275": "lol! What you said cracks me up. Yeah, most of the code right now is model ensembling. You can check out the code titled train. Currently, the ByT5 model is what most people are using, but you can give other models a try.",
    "3394427": "The models shared only serve as the baseline -- they could be used for further finetuning, ensemble, or just for testing a preprocess idea. However, to succeed in the competition you need more than rerun these inference notebooks."
  },
  "source": "meta"
}