{
  "id": 671317,
  "title": "current RNAPro performance",
  "url": "/competitions/stanford-rna-3d-folding-2/discussion/671317",
  "author_name": "Kh0a",
  "post_date": "2026-02-01T08:41:14.663000",
  "votes": 8,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I haven't tried fine-tuning RNAPro or other models yet, but based on public kernels' LB scores, I observed:</p>\n<ul>\n<li><p>Pure TBM reached the highest LB score: <a href=\"https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108\" target=\"_blank\">https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108</a> — 0.366 </p></li>\n<li><p>Best LB score for a notebook that actually used both TBM and RNAPro: <a href=\"https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook\" target=\"_blank\">https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook</a> — 0.357</p></li>\n</ul>\n<p>It appears the current RNAPro setup may be hurting the TBM approach. I believe RNAPro has potential to further enhance TBM, but I'm unsure about the best strategy:</p>\n<p>Should I train RNAPro using only the dataset from this competition, or fine-tune it from the existing RNAPro checkpoint (which was trained on the previous Stanford RNA 3D Folding dataset)? </p>\n<p>What are your thoughts on this?\nThanks</p>",
  "messages": [
    {
      "id": 3400252,
      "postDate": "2026-02-01T08:41:14.663Z",
      "content": "<p>I haven't tried fine-tuning RNAPro or other models yet, but based on public kernels' LB scores, I observed:</p>\n<ul>\n<li><p>Pure TBM reached the highest LB score: <a href=\"https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108\" target=\"_blank\">https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108</a> — 0.366 </p></li>\n<li><p>Best LB score for a notebook that actually used both TBM and RNAPro: <a href=\"https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook\" target=\"_blank\">https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook</a> — 0.357</p></li>\n</ul>\n<p>It appears the current RNAPro setup may be hurting the TBM approach. I believe RNAPro has potential to further enhance TBM, but I'm unsure about the best strategy:</p>\n<p>Should I train RNAPro using only the dataset from this competition, or fine-tune it from the existing RNAPro checkpoint (which was trained on the previous Stanford RNA 3D Folding dataset)? </p>\n<p>What are your thoughts on this?\nThanks</p>",
      "rawMarkdown": "I haven't tried fine-tuning RNAPro or other models yet, but based on public kernels' LB scores, I observed:\n\n- Pure TBM reached the highest LB score: https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108 — 0.366 \n\n\n- Best LB score for a notebook that actually used both TBM and RNAPro: https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook — 0.357\n\n\nIt appears the current RNAPro setup may be hurting the TBM approach. I believe RNAPro has potential to further enhance TBM, but I'm unsure about the best strategy:\n\nShould I train RNAPro using only the dataset from this competition, or fine-tune it from the existing RNAPro checkpoint (which was trained on the previous Stanford RNA 3D Folding dataset)? \n\nWhat are your thoughts on this?\nThanks",
      "votes": 7
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3400252": "I haven't tried fine-tuning RNAPro or other models yet, but based on public kernels' LB scores, I observed:\n\n- Pure TBM reached the highest LB score: https://www.kaggle.com/code/jakupymeraj/rna-folding-prediction-31jan?scriptVersionId=295071108 — 0.366 \n\n\n- Best LB score for a notebook that actually used both TBM and RNAPro: https://www.kaggle.com/code/andrewlee23023/rnapro-inference-with-tbm/notebook — 0.357\n\n\nIt appears the current RNAPro setup may be hurting the TBM approach. I believe RNAPro has potential to further enhance TBM, but I'm unsure about the best strategy:\n\nShould I train RNAPro using only the dataset from this competition, or fine-tune it from the existing RNAPro checkpoint (which was trained on the previous Stanford RNA 3D Folding dataset)? \n\nWhat are your thoughts on this?\nThanks"
  }
}