Dear Host, @rhijudas @przemekporebski — could you please clarify two points about Stanford RNA 3D Folding Part 2?
\n(Apologies if this has already been answered somewhere — feel free to point me to the relevant post/link , thanks.)
\nMany Kaggle competitions explicitly state how the public leaderboard is computed (e.g., “~20% public / ~80% private”), but I don’t see that information on the competition page in this comp:\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding-2
\nWill the final results be determined by re-running/re-scoring submissions after the competition ends (i.e., a standardized rerun), or will scoring be based only on the submitted outputs as-is?
\nI’m asking because some inference pipelines (e.g., Protenix-based) can introduce noticeable randomness, so understanding the final evaluation procedure is important for reproducibility and final standings.
\nThanks in advance for the clarification.
", "messages": [ { "id": 3416259, "postDate": "2026-03-02T13:06:51.897Z", "content": "Dear Host, @rhijudas @przemekporebski — could you please clarify two points about Stanford RNA 3D Folding Part 2?
\n(Apologies if this has already been answered somewhere — feel free to point me to the relevant post/link , thanks.)
\nMany Kaggle competitions explicitly state how the public leaderboard is computed (e.g., “~20% public / ~80% private”), but I don’t see that information on the competition page in this comp:\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding-2
\nWill the final results be determined by re-running/re-scoring submissions after the competition ends (i.e., a standardized rerun), or will scoring be based only on the submitted outputs as-is?
\nI’m asking because some inference pipelines (e.g., Protenix-based) can introduce noticeable randomness, so understanding the final evaluation procedure is important for reproducibility and final standings.
\nThanks in advance for the clarification.
", "rawMarkdown": "Dear Host, @rhijudas @przemekporebski — could you please clarify two points about **Stanford RNA 3D Folding Part 2**?\n\n*(Apologies if this has already been answered somewhere — feel free to point me to the relevant post/link , thanks.)*\n\n### 1) Public vs. private leaderboard split (test data distribution)\n\nMany Kaggle competitions explicitly state how the public leaderboard is computed (e.g., “~20% public / ~80% private”), but I don’t see that information on the competition page in this comp:\nhttps://www.kaggle.com/competitions/stanford-rna-3d-folding-2 \n\n* Is the public leaderboard based on a fixed subset of the test set?\n* If so, what is the approximate public/private percentage split?\n\n### 2) Final scoring / re-run policy\n\nWill the final results be determined by **re-running/re-scoring submissions** after the competition ends (i.e., a standardized rerun), or will scoring be based only on the submitted outputs as-is?\n\nI’m asking because some inference pipelines (e.g., Protenix-based) can introduce noticeable randomness, so understanding the final evaluation procedure is important for reproducibility and final standings.\n\nThanks in advance for the clarification.\n", "votes": 9 }, { "id": 3416707, "postDate": "2026-03-03T14:18:47.310Z", "content": "Good questions -- thanks for asking them and congratulations on your progress so far!
\n1) Public vs. private leaderboard split
\nTo help reduce chances of leakage, experimental data for the private leaderboard targets are being collected during the competition. Since we don't know which RNA targets we'll get high quality data for, we can't provide a percentage estimate of public/private split until the competition closes.
\n2) Final scoring / re-run policy
\nThe final scores will be based on re-run of the two notebooks that each team selects. For notebooks that use random seeds, there will indeed be some element of chance. These notebooks can reduce random fluctuations by using fixed seeds or selecting models from multiple runs.
\nGood luck!
", "rawMarkdown": "Good questions -- thanks for asking them and congratulations on your progress so far!\n\n**1) Public vs. private leaderboard split**\n\nTo help reduce chances of leakage, experimental data for the private leaderboard targets are being collected *during* the competition. Since we don't know which RNA targets we'll get high quality data for, we can't provide a percentage estimate of public/private split until the competition closes.\n\n**2) Final scoring / re-run policy**\n\nThe final scores will be based on re-run of the two notebooks that each team selects. For notebooks that use random seeds, there will indeed be some element of chance. These notebooks can reduce random fluctuations by using fixed seeds or selecting models from multiple runs.\n\nGood luck! ", "votes": 2, "replies": [ { "id": 3416747, "postDate": "2026-03-03T16:20:04.147Z", "content": "Dear host @rhijudas but there is time constraint that the notebook should be <= 8hours of gpu or cpu if you can say the number of rna sequence tested it can be good to optimize the inference based on that . Thank you.
", "rawMarkdown": "Dear host @rhijudas but there is time constraint that the notebook should be <= 8hours of gpu or cpu if you can say the number of rna sequence tested it can be good to optimize the inference based on that . Thank you.", "votes": 2, "replies": [ { "id": 3416751, "postDate": "2026-03-03T16:29:35.217Z", "content": "That’s a good question.
\nAlthough we can’t reveal number of sequences for the rerun, it will be similar to the number of sequences that notebooks are being run on now.
\nSo notebooks that run and score properly now in the training phase should be fine in the rerun.
\nNote that we did something similar in Part 1 of this competition and the vast majority of notebooks were rerun successfully - it was great to see that Kagglers engineered their codes for robustness, and we hope that continues!
", "rawMarkdown": "That’s a good question. \n\nAlthough we can’t reveal number of sequences for the rerun, it will be similar to the number of sequences that notebooks are being run on now. \n\nSo notebooks that run and score properly now in the training phase should be fine in the rerun. \n\nNote that we did something similar in Part 1 of this competition and the vast majority of notebooks were rerun successfully - it was great to see that Kagglers engineered their codes for robustness, and we hope that continues!", "votes": 2 }, { "id": 3420751, "postDate": "2026-03-13T21:41:13.110Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3427254, "postDate": "2026-03-23T21:45:50.093Z", "content": "Will you maintain the hard 8 hour limit? The competition description says the time may be expanded. It's a lot easier to engineer for 8 hrs compared with an expansion based on the number of items, because we don't know how many items or their lengths now.
", "rawMarkdown": "Will you maintain the hard 8 hour limit? The competition description says the time may be expanded. It's a lot easier to engineer for 8 hrs compared with an expansion based on the number of items, because we don't know how many items or their lengths now.", "votes": 1 }, { "id": 3427577, "postDate": "2026-03-24T09:06:17.353Z", "content": "It would be nice if Kaggle could provide the notebook runtime limit in a variable, in the same spirit as KAGGLE_IS_COMPETITION_RERUN. Maybe it could be a useful feature for competition hosts…
Good questions -- thanks for asking them and congratulations on your progress so far!
\n1) Public vs. private leaderboard split
\nTo help reduce chances of leakage, experimental data for the private leaderboard targets are being collected during the competition. Since we don't know which RNA targets we'll get high quality data for, we can't provide a percentage estimate of public/private split until the competition closes.
\n2) Final scoring / re-run policy
\nThe final scores will be based on re-run of the two notebooks that each team selects. For notebooks that use random seeds, there will indeed be some element of chance. These notebooks can reduce random fluctuations by using fixed seeds or selecting models from multiple runs.
\nGood luck!
", "votes": 2, "replies": [ { "id": 3416747, "author_name": "Ragesh Thangaraj", "author_url": "", "post_date": "2026-03-03T16:20:04.147000", "content": "Dear host @rhijudas but there is time constraint that the notebook should be <= 8hours of gpu or cpu if you can say the number of rna sequence tested it can be good to optimize the inference based on that . Thank you.
", "votes": 2, "replies": [ { "id": 3416751, "author_name": "Rhiju Das", "author_url": "", "post_date": "2026-03-03T16:29:35.217000", "content": "That’s a good question.
\nAlthough we can’t reveal number of sequences for the rerun, it will be similar to the number of sequences that notebooks are being run on now.
\nSo notebooks that run and score properly now in the training phase should be fine in the rerun.
\nNote that we did something similar in Part 1 of this competition and the vast majority of notebooks were rerun successfully - it was great to see that Kagglers engineered their codes for robustness, and we hope that continues!
", "votes": 2, "replies": [] }, { "id": 3420751, "author_name": "", "author_url": "", "post_date": "2026-03-13T21:41:13.110000", "content": "", "votes": 0, "replies": [] }, { "id": 3427254, "author_name": "Jack Cole", "author_url": "", "post_date": "2026-03-23T21:45:50.093000", "content": "Will you maintain the hard 8 hour limit? The competition description says the time may be expanded. It's a lot easier to engineer for 8 hrs compared with an expansion based on the number of items, because we don't know how many items or their lengths now.
", "votes": 1, "replies": [] }, { "id": 3427577, "author_name": "Gabor Balazs", "author_url": "", "post_date": "2026-03-24T09:06:17.353000", "content": "It would be nice if Kaggle could provide the notebook runtime limit in a variable, in the same spirit as KAGGLE_IS_COMPETITION_RERUN. Maybe it could be a useful feature for competition hosts…