{
  "id": 668783,
  "title": "Do we need to train the backbone of the DINO model? ",
  "url": "/competitions/csiro-biomass/discussion/668783",
  "author_name": "",
  "post_date": "2026-01-19T02:18:32.187937600Z",
  "votes": -2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I've seen many open-source code solutions freeze the backbone, so would training the backbone lead to an improvement in the LB score?</p>",
  "messages": [
    {
      "id": "3393421",
      "postDate": "01/19/2026 02:18:32",
      "content": "<p>I've seen many open-source code solutions freeze the backbone, so would training the backbone lead to an improvement in the LB score?</p>",
      "rawMarkdown": "I've seen many open-source code solutions freeze the backbone, so would training the backbone lead to an improvement in the LB score?",
      "votes": null
    },
    {
      "id": "3394180",
      "postDate": "01/20/2026 15:34:51",
      "content": "<p>you know you can just try things and find out …</p>",
      "rawMarkdown": "you know you can just try things and find out …",
      "votes": null
    },
    {
      "id": "3394248",
      "postDate": "01/20/2026 17:45:47",
      "content": "<p>dataset's too small to unfreeze huge models like the dino models. you'll overfit instantly. try finetuning with LoRA or block expansion. I'm still figuring out the right folds, but be careful not to chase the public lb score.</p>",
      "rawMarkdown": "dataset's too small to unfreeze huge models like the dino models. you'll overfit instantly. try finetuning with LoRA or block expansion. I'm still figuring out the right folds, but be careful not to chase the public lb score.",
      "votes": null
    },
    {
      "id": "3394278",
      "postDate": "01/20/2026 19:10:00",
      "content": "<p>I have found a slight improvement with small percentage range of freezes.   BUT the dataset is so small, and the test set is not much bigger making both CV and LB score pretty shaky indicators of real or false improvement.</p>\n<p>This is a great competition to learn some small data set sample tools and skills.</p>",
      "rawMarkdown": "I have found a slight improvement with small percentage range of freezes.   BUT the dataset is so small, and the test set is not much bigger making both CV and LB score pretty shaky indicators of real or false improvement.\n\nThis is a great competition to learn some small data set sample tools and skills.",
      "votes": null
    },
    {
      "id": "3395333",
      "postDate": "01/22/2026 18:38:58",
      "content": "<p>Up till now, I was unfreezing the backbone. But I recently changed some gears and now I am able to get identical performance without touching the backbone at all. So I would say, stick with unfreezing for now while you figure out all the rest of the things. Keep the idea for later. But note that unfreezing won't make a huge difference 'here'. If it does, understand that you are going on the wrong track.</p>",
      "rawMarkdown": "Up till now, I was unfreezing the backbone. But I recently changed some gears and now I am able to get identical performance without touching the backbone at all. So I would say, stick with unfreezing for now while you figure out all the rest of the things. Keep the idea for later. But note that unfreezing won't make a huge difference 'here'. If it does, understand that you are going on the wrong track.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3394180,
      "author_name": "yeoyunsianggeremie",
      "author_url": "",
      "post_date": "01/20/2026 15:34:51",
      "content": "<p>you know you can just try things and find out …</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3394248,
      "author_name": "eddieiyamah",
      "author_url": "",
      "post_date": "01/20/2026 17:45:47",
      "content": "<p>dataset's too small to unfreeze huge models like the dino models. you'll overfit instantly. try finetuning with LoRA or block expansion. I'm still figuring out the right folds, but be careful not to chase the public lb score.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3394278,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "01/20/2026 19:10:00",
      "content": "<p>I have found a slight improvement with small percentage range of freezes.   BUT the dataset is so small, and the test set is not much bigger making both CV and LB score pretty shaky indicators of real or false improvement.</p>\n<p>This is a great competition to learn some small data set sample tools and skills.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3395333,
      "author_name": "komilparmar",
      "author_url": "",
      "post_date": "01/22/2026 18:38:58",
      "content": "<p>Up till now, I was unfreezing the backbone. But I recently changed some gears and now I am able to get identical performance without touching the backbone at all. So I would say, stick with unfreezing for now while you figure out all the rest of the things. Keep the idea for later. But note that unfreezing won't make a huge difference 'here'. If it does, understand that you are going on the wrong track.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3393421": "I've seen many open-source code solutions freeze the backbone, so would training the backbone lead to an improvement in the LB score?",
    "3394180": "you know you can just try things and find out …",
    "3394248": "dataset's too small to unfreeze huge models like the dino models. you'll overfit instantly. try finetuning with LoRA or block expansion. I'm still figuring out the right folds, but be careful not to chase the public lb score.",
    "3394278": "I have found a slight improvement with small percentage range of freezes.   BUT the dataset is so small, and the test set is not much bigger making both CV and LB score pretty shaky indicators of real or false improvement.\n\nThis is a great competition to learn some small data set sample tools and skills.",
    "3395333": "Up till now, I was unfreezing the backbone. But I recently changed some gears and now I am able to get identical performance without touching the backbone at all. So I would say, stick with unfreezing for now while you figure out all the rest of the things. Keep the idea for later. But note that unfreezing won't make a huge difference 'here'. If it does, understand that you are going on the wrong track."
  },
  "source": "meta"
}