{
  "id": 570169,
  "title": "Does it make sense to train using the training set?",
  "url": "/competitions/stanford-rna-3d-folding/discussion/570169",
  "author_name": "doheon114",
  "post_date": "2025-03-26T08:12:39.230000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>With only about 800 RNA sequences, would it be effective to fine-tune an open-source model designed for the overall prediction of proteins, DNA, RNA, and ligands to ultimately improve the TM-score?</p>\n<p>To the best of my knowledge, fine-tuning is typically useful for domain adaptation, such as developing disease-specific or institution-specific models. However, I am skeptical about its efficacy in this case because RNA is a physical entity. I believe fine-tuning assumes that the training set and the hidden test set share a similar structural distribution, which may not always hold true.</p>",
  "messages": [
    {
      "id": 3160013,
      "postDate": "2025-03-26T08:12:39.230Z",
      "content": "<p>With only about 800 RNA sequences, would it be effective to fine-tune an open-source model designed for the overall prediction of proteins, DNA, RNA, and ligands to ultimately improve the TM-score?</p>\n<p>To the best of my knowledge, fine-tuning is typically useful for domain adaptation, such as developing disease-specific or institution-specific models. However, I am skeptical about its efficacy in this case because RNA is a physical entity. I believe fine-tuning assumes that the training set and the hidden test set share a similar structural distribution, which may not always hold true.</p>",
      "rawMarkdown": "With only about 800 RNA sequences, would it be effective to fine-tune an open-source model designed for the overall prediction of proteins, DNA, RNA, and ligands to ultimately improve the TM-score?\n\nTo the best of my knowledge, fine-tuning is typically useful for domain adaptation, such as developing disease-specific or institution-specific models. However, I am skeptical about its efficacy in this case because RNA is a physical entity. I believe fine-tuning assumes that the training set and the hidden test set share a similar structural distribution, which may not always hold true.",
      "votes": 1
    },
    {
      "id": 3160090,
      "postDate": "2025-03-26T11:01:19.147Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3160090,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-03-26T11:01:19.147000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3160013": "With only about 800 RNA sequences, would it be effective to fine-tune an open-source model designed for the overall prediction of proteins, DNA, RNA, and ligands to ultimately improve the TM-score?\n\nTo the best of my knowledge, fine-tuning is typically useful for domain adaptation, such as developing disease-specific or institution-specific models. However, I am skeptical about its efficacy in this case because RNA is a physical entity. I believe fine-tuning assumes that the training set and the hidden test set share a similar structural distribution, which may not always hold true.",
    "3160090": ""
  }
}