Despite the 53% of data used for public leaderboard, I expect scores on the private leaderboard to be different. My best guess is that the winning team will not exceed 0.72…, so less than 0.73.
\nThe reason is that biomass targets seem to have a very different meaning for different predictive approaches.
\nWhat do you think the range of scores will be?
\nDespite the 53% of data used for public leaderboard, I expect scores on the private leaderboard to be different. My best guess is that the winning team will not exceed 0.72…, so less than 0.73.
\nThe reason is that biomass targets seem to have a very different meaning for different predictive approaches.
\nWhat do you think the range of scores will be?
\nThe current highest score is 0.8. From 0.8 to 0.72 would be a huge jump. I don't believe that would happen. The CV scores atleast in my case are very very consistent. I am always getting exactly what I am expecting on LB. I hence think, the winner will have a score at least as high as 0.77. But more importantly, the leaderboard is barely going to change in terms of ranks, and that shakeups too are going to be just because of blind trial and error and selecting the trial which simply performs the best, removing some folds to improve the public LB score etc.
", "rawMarkdown": "The current highest score is 0.8. From 0.8 to 0.72 would be a huge jump. I don't believe that would happen. The CV scores atleast in my case are very very consistent. I am always getting exactly what I am expecting on LB. I hence think, the winner will have a score at least as high as 0.77. But more importantly, the leaderboard is barely going to change in terms of ranks, and that shakeups too are going to be just because of blind trial and error and selecting the trial which simply performs the best, removing some folds to improve the public LB score etc.", "votes": null }, { "id": "3397360", "postDate": "01/27/2026 05:07:42", "content": "i also think most shakeups would be for people just forking public notebooks with blind post processing and the other sort of things involved to overfit the leaderboard otherwise there wont be much change
", "rawMarkdown": "i also think most shakeups would be for people just forking public notebooks with blind post processing and the other sort of things involved to overfit the leaderboard otherwise there wont be much change", "votes": null }, { "id": "3397392", "postDate": "01/27/2026 07:15:02", "content": "Given the early behavior of the leaderboard with respect to local CV and the fact that the training data is very small, I believe many top teams will drop, though I do not wish it for them. It is just inevitable. No high score will be seen on the private leaderboard. We may see new top teams with signs like +648, +1024, etc. All the best to all teams indeed.
", "rawMarkdown": "Given the early behavior of the leaderboard with respect to local CV and the fact that the training data is very small, I believe many top teams will drop, though I do not wish it for them. It is just inevitable. No high score will be seen on the private leaderboard. We may see new top teams with signs like +648, +1024, etc. All the best to all teams indeed.", "votes": null }, { "id": "3397398", "postDate": "01/27/2026 07:27:00", "content": "I think the possibility of this is not high.
", "rawMarkdown": "I think the possibility of this is not high.", "votes": null }, { "id": "3397419", "postDate": "01/27/2026 08:21:56", "content": "Rather than just looking into ups and downs, we need to see why these could happen @jeannkouagou \nWe have the below possibilities -
\n| Cause | \nImpact | \n
|---|---|
| Teams blindly relying on luck on the public leaderboard with questionable model selection | \nHigh chance of shakeup | \n
| Teams relying on questionable design choices and scalers | \nModerate chance of shakeup - can't comment on post-processing | \n
| Teams relying on sound design choices and models | \nLow chance of shakeup | \n
I think we will have localized shakeups here, example -
I wish anyone anyone can help me answer this question. The answer could either show I am wrong or right.
\nI think the host split the dataset by year. If you look closely, you’ll see that the entire training dataset is from 2015, missing only a single month, while the data was collected over three years.
\nAnother point is that the training dataset has approximately 370 samples (I don’t remember the exact number), and the test dataset has a total of 800 samples. So we can guess that each year has roughly 400 samples. That’s why the ratio of public to private is approximately 53:47.
", "rawMarkdown": "I think the host split the dataset by year. If you look closely, you’ll see that the entire training dataset is from 2015, missing only a single month, while the data was collected over three years.\n\nAnother point is that the training dataset has approximately 370 samples (I don’t remember the exact number), and the test dataset has a total of 800 samples. So we can guess that each year has roughly 400 samples. That’s why the ratio of public to private is approximately 53:47.", "votes": null }, { "id": "3397456", "postDate": "01/27/2026 09:39:39", "content": "I agree that would explain the weird split and could also be a source of shakeup. Although there is probably a clever way to use it too.
", "rawMarkdown": "I agree that would explain the weird split and could also be a source of shakeup. Although there is probably a clever way to use it too.", "votes": null }, { "id": "3397478", "postDate": "01/27/2026 10:51:17", "content": "Hi @jeannkouagou , I am curious why do you guess the number of 0.72 or 0.73?
", "rawMarkdown": "Hi @jeannkouagou , I am curious why do you guess the number of 0.72 or 0.73?", "votes": null }, { "id": "3397611", "postDate": "01/27/2026 17:17:09", "content": "I hope this to be true i have been trying to push beyond bbut stuck at 0.73,kinda relieved after seeing your comment @jeannkouagou \nTHANK YOU
", "rawMarkdown": "I hope this to be true i have been trying to push beyond bbut stuck at 0.73,kinda relieved after seeing your comment @jeannkouagou \nTHANK YOU", "votes": null }, { "id": "3397613", "postDate": "01/27/2026 17:18:36", "content": "@forcewithme this is a cutoff for a lot of people where the cv-lb relation starts to turn awry
", "rawMarkdown": "forcewithme this is a cutoff for a lot of people where the cv-lb relation starts to turn awry", "votes": null }, { "id": "3397616", "postDate": "01/27/2026 17:22:22", "content": "@komilparmar I think the private leaderboard score is likely to be similar to the public leaderboard score with scores ranging in the higher 0.70s for winning solutions and 0.70/ 0.69 for bronze closures.
\nThe dispersion between the highest ranked submission (0.81) to the last bronze ranked submission as on date (0.72) is quite widespread and unusual for Kaggle at least!
", "rawMarkdown": "komilparmar I think the private leaderboard score is likely to be similar to the public leaderboard score with scores ranging in the higher 0.70s for winning solutions and 0.70/ 0.69 for bronze closures. \n\nThe dispersion between the highest ranked submission (0.81) to the last bronze ranked submission as on date (0.72) is quite widespread and unusual for Kaggle at least!", "votes": null }, { "id": "3398152", "postDate": "01/28/2026 17:42:24", "content": "0.72 could even be a high estimate for the private leaderboard. If the private test set comes from a different year, we may have completely different biomass targets. But I might be wrong and maybe participants have developed very strong models that can generalize to any test set distribution, like a model that can really \"see\" even better than human experts.
", "rawMarkdown": "0.72 could even be a high estimate for the private leaderboard. If the private test set comes from a different year, we may have completely different biomass targets. But I might be wrong and maybe participants have developed very strong models that can generalize to any test set distribution, like a model that can really \"see\" even better than human experts.", "votes": null }, { "id": "3398165", "postDate": "01/28/2026 18:01:40", "content": "The only 3 things that comes to my mind seeing a score higher than 0.74 are
\nThe beauty of Pseudo labelling especially is that, the model never discriminates between samples of public and private lbs. Because it's doesn't know how to do that. I believe if this technique or other such test data based post processing is what is getting these people so high scores, they will indeed more or less get the same scores or atleast near same ranks on private lb.
", "rawMarkdown": "The only 3 things that comes to my mind seeing a score higher than 0.74 are\n1. Pretraining backbines on Domain\n2. Using external labelled datasets\n3. Supreme Post processing or Pseudo labelling. \n\nThe beauty of Pseudo labelling especially is that, the model never discriminates between samples of public and private lbs. Because it's doesn't know how to do that. I believe if this technique or other such test data based post processing is what is getting these people so high scores, they will indeed more or less get the same scores or atleast near same ranks on private lb.", "votes": null }, { "id": "3398168", "postDate": "01/28/2026 18:13:20", "content": "I have done 2 and 3 without much success. But maybe some others did it better. Pseudo labelling works to an extent but there is a tradeoff because of the compute constraints. There are a lot of unlabeled external images that could be pseudo labeled too (haven't tried that).
", "rawMarkdown": "I have done 2 and 3 without much success. But maybe some others did it better. Pseudo labelling works to an extent but there is a tradeoff because of the compute constraints. There are a lot of unlabeled external images that could be pseudo labeled too (haven't tried that).", "votes": null }, { "id": "3398192", "postDate": "01/28/2026 19:01:59", "content": "We have got only 5h to see the results, so excited 😆
", "rawMarkdown": "We have got only 5h to see the results, so excited 😆", "votes": null }, { "id": "3398509", "postDate": "01/29/2026 08:39:04", "content": "So the private line leaderboard just confirmed my score range. Congrats to all the winners, see you in another comp
", "rawMarkdown": "So the private line leaderboard just confirmed my score range. Congrats to all the winners, see you in another comp", "votes": null } ], "comments": [ { "id": 3397312, "author_name": "komilparmar", "author_url": "", "post_date": "01/27/2026 01:12:03", "content": "The current highest score is 0.8. From 0.8 to 0.72 would be a huge jump. I don't believe that would happen. The CV scores atleast in my case are very very consistent. I am always getting exactly what I am expecting on LB. I hence think, the winner will have a score at least as high as 0.77. But more importantly, the leaderboard is barely going to change in terms of ranks, and that shakeups too are going to be just because of blind trial and error and selecting the trial which simply performs the best, removing some folds to improve the public LB score etc.
", "votes": null, "replies": [ { "id": 3397360, "author_name": "suman2208", "author_url": "", "post_date": "01/27/2026 05:07:42", "content": "i also think most shakeups would be for people just forking public notebooks with blind post processing and the other sort of things involved to overfit the leaderboard otherwise there wont be much change
", "votes": null, "replies": [] }, { "id": 3397616, "author_name": "ravi20076", "author_url": "", "post_date": "01/27/2026 17:22:22", "content": "@komilparmar I think the private leaderboard score is likely to be similar to the public leaderboard score with scores ranging in the higher 0.70s for winning solutions and 0.70/ 0.69 for bronze closures.
\nThe dispersion between the highest ranked submission (0.81) to the last bronze ranked submission as on date (0.72) is quite widespread and unusual for Kaggle at least!
", "votes": null, "replies": [] } ] }, { "id": 3397392, "author_name": "jeannkouagou", "author_url": "", "post_date": "01/27/2026 07:15:02", "content": "Given the early behavior of the leaderboard with respect to local CV and the fact that the training data is very small, I believe many top teams will drop, though I do not wish it for them. It is just inevitable. No high score will be seen on the private leaderboard. We may see new top teams with signs like +648, +1024, etc. All the best to all teams indeed.
", "votes": null, "replies": [ { "id": 3397398, "author_name": "kagger3", "author_url": "", "post_date": "01/27/2026 07:27:00", "content": "I think the possibility of this is not high.
", "votes": null, "replies": [] }, { "id": 3397419, "author_name": "ravi20076", "author_url": "", "post_date": "01/27/2026 08:21:56", "content": "Rather than just looking into ups and downs, we need to see why these could happen @jeannkouagou \nWe have the below possibilities -
\n| Cause | \nImpact | \n
|---|---|
| Teams blindly relying on luck on the public leaderboard with questionable model selection | \nHigh chance of shakeup | \n
| Teams relying on questionable design choices and scalers | \nModerate chance of shakeup - can't comment on post-processing | \n
| Teams relying on sound design choices and models | \nLow chance of shakeup | \n
I think we will have localized shakeups here, example -
I wish anyone anyone can help me answer this question. The answer could either show I am wrong or right.
\nI think the host split the dataset by year. If you look closely, you’ll see that the entire training dataset is from 2015, missing only a single month, while the data was collected over three years.
\nAnother point is that the training dataset has approximately 370 samples (I don’t remember the exact number), and the test dataset has a total of 800 samples. So we can guess that each year has roughly 400 samples. That’s why the ratio of public to private is approximately 53:47.
", "votes": null, "replies": [ { "id": 3397456, "author_name": "tennogh", "author_url": "", "post_date": "01/27/2026 09:39:39", "content": "I agree that would explain the weird split and could also be a source of shakeup. Although there is probably a clever way to use it too.
", "votes": null, "replies": [] } ] } ] }, { "id": 3397478, "author_name": "forcewithme", "author_url": "", "post_date": "01/27/2026 10:51:17", "content": "Hi @jeannkouagou , I am curious why do you guess the number of 0.72 or 0.73?
", "votes": null, "replies": [ { "id": 3397613, "author_name": "ravi20076", "author_url": "", "post_date": "01/27/2026 17:18:36", "content": "@forcewithme this is a cutoff for a lot of people where the cv-lb relation starts to turn awry
", "votes": null, "replies": [] }, { "id": 3398152, "author_name": "jeannkouagou", "author_url": "", "post_date": "01/28/2026 17:42:24", "content": "0.72 could even be a high estimate for the private leaderboard. If the private test set comes from a different year, we may have completely different biomass targets. But I might be wrong and maybe participants have developed very strong models that can generalize to any test set distribution, like a model that can really \"see\" even better than human experts.
", "votes": null, "replies": [ { "id": 3398165, "author_name": "komilparmar", "author_url": "", "post_date": "01/28/2026 18:01:40", "content": "The only 3 things that comes to my mind seeing a score higher than 0.74 are
\nThe beauty of Pseudo labelling especially is that, the model never discriminates between samples of public and private lbs. Because it's doesn't know how to do that. I believe if this technique or other such test data based post processing is what is getting these people so high scores, they will indeed more or less get the same scores or atleast near same ranks on private lb.
", "votes": null, "replies": [ { "id": 3398168, "author_name": "tennogh", "author_url": "", "post_date": "01/28/2026 18:13:20", "content": "I have done 2 and 3 without much success. But maybe some others did it better. Pseudo labelling works to an extent but there is a tradeoff because of the compute constraints. There are a lot of unlabeled external images that could be pseudo labeled too (haven't tried that).
", "votes": null, "replies": [ { "id": 3398192, "author_name": "jeannkouagou", "author_url": "", "post_date": "01/28/2026 19:01:59", "content": "We have got only 5h to see the results, so excited 😆
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3397611, "author_name": "sachchidanandadaki", "author_url": "", "post_date": "01/27/2026 17:17:09", "content": "I hope this to be true i have been trying to push beyond bbut stuck at 0.73,kinda relieved after seeing your comment @jeannkouagou \nTHANK YOU
", "votes": null, "replies": [] }, { "id": 3398509, "author_name": "jeannkouagou", "author_url": "", "post_date": "01/29/2026 08:39:04", "content": "So the private line leaderboard just confirmed my score range. Congrats to all the winners, see you in another comp
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3397252": "Despite the 53% of data used for public leaderboard, I expect scores on the private leaderboard to be different. My best guess is that the winning team will not exceed 0.72..., so less than 0.73.\n\nThe reason is that biomass targets seem to have a very different meaning for different predictive approaches. \n\nWhat do you think the range of scores will be?\n\n## Good Luck!", "3397312": "The current highest score is 0.8. From 0.8 to 0.72 would be a huge jump. I don't believe that would happen. The CV scores atleast in my case are very very consistent. I am always getting exactly what I am expecting on LB. I hence think, the winner will have a score at least as high as 0.77. But more importantly, the leaderboard is barely going to change in terms of ranks, and that shakeups too are going to be just because of blind trial and error and selecting the trial which simply performs the best, removing some folds to improve the public LB score etc.", "3397360": "i also think most shakeups would be for people just forking public notebooks with blind post processing and the other sort of things involved to overfit the leaderboard otherwise there wont be much change", "3397392": "Given the early behavior of the leaderboard with respect to local CV and the fact that the training data is very small, I believe many top teams will drop, though I do not wish it for them. It is just inevitable. No high score will be seen on the private leaderboard. We may see new top teams with signs like +648, +1024, etc. All the best to all teams indeed.", "3397398": "I think the possibility of this is not high.", "3397419": "Rather than just looking into ups and downs, we need to see why these could happen @jeannkouagou \nWe have the below possibilities - \n\n|Cause|Impact|\n|----- |------|\n|Teams blindly relying on luck on the public leaderboard with questionable model selection | High chance of shakeup |\n|Teams relying on questionable design choices and scalers | Moderate chance of shakeup - can't comment on post-processing |\n|Teams relying on sound design choices and models | Low chance of shakeup |\n\n