Just a simple request .. I know i will get downvote for this but only few days are left in the competition and people are just posting the higher scoring submission notebook… It really effect the nature of the competition as many people are trying to score good by pushing hard on daily basis .. don't ruin this fun for everyone , it doesn't look good :)
\nPS : Currently at 210th place and slipping on daily basis lol
", "messages": [ { "id": 3393066, "postDate": "2026-01-18T07:22:10.193Z", "content": "Just a simple request .. I know i will get downvote for this but only few days are left in the competition and people are just posting the higher scoring submission notebook… It really effect the nature of the competition as many people are trying to score good by pushing hard on daily basis .. don't ruin this fun for everyone , it doesn't look good :)
\nPS : Currently at 210th place and slipping on daily basis lol
", "rawMarkdown": "Just a simple request .. I know i will get downvote for this but only few days are left in the competition and people are just posting the higher scoring submission notebook... It really effect the nature of the competition as many people are trying to score good by pushing hard on daily basis .. don't ruin this fun for everyone , it doesn't look good :)\n\nPS : Currently at 210th place and slipping on daily basis lol ", "votes": 41 }, { "id": 3394757, "postDate": "2026-01-21T16:23:00.217Z", "content": "\"I totally agree with you! The 'Late Submission Dumping' takes away the real learning spirit of the competition. It's frustrating to see hard-earned ranks slip due to public notebooks. Keep pushing, being in the top 250 by your own effort is much more valuable for your AGRO_TECH skills than just clicking 'run' on a public script! Upvoted for visibility.\"
", "rawMarkdown": "\"I totally agree with you! The 'Late Submission Dumping' takes away the real learning spirit of the competition. It's frustrating to see hard-earned ranks slip due to public notebooks. Keep pushing, being in the top 250 by your own effort is much more valuable for your AGRO_TECH skills than just clicking 'run' on a public script! Upvoted for visibility.\"", "votes": 4 }, { "id": 3393083, "postDate": "2026-01-18T08:00:38.123Z", "content": "Kaggle competitions are becoming increasingly difficult; in this competition, a shake-up is likely, and the private score of teams in the public gold zone could end up lower than the private score of a public notebook 🥱🥱🥱
", "rawMarkdown": "Kaggle competitions are becoming increasingly difficult; in this competition, a shake-up is likely, and the private score of teams in the public gold zone could end up lower than the private score of a public notebook 🥱🥱🥱", "votes": 2, "replies": [ { "id": 3393094, "postDate": "2026-01-18T08:11:21.123Z", "content": "How sure are you there would be a shake?
", "rawMarkdown": "How sure are you there would be a shake?", "replies": [ { "id": 3393163, "postDate": "2026-01-18T11:00:52.443Z", "content": "one reason could be using scaling factor i think, as the scaling factor would only for now , not when whole test set is used causing wrong total and etc
", "rawMarkdown": "one reason could be using scaling factor i think, as the scaling factor would only for now , not when whole test set is used causing wrong total and etc\n" }, { "id": 3393164, "postDate": "2026-01-18T11:03:32.387Z", "content": "There is too little data, while the value ranges for dead, green, and clover are very wide. Meanwhile, the host split the data by year, and depending on the year, the value distribution of biomass can be different. Sometimes a simple solution might actually be better. Luck is all we need.
", "rawMarkdown": "There is too little data, while the value ranges for dead, green, and clover are very wide. Meanwhile, the host split the data by year, and depending on the year, the value distribution of biomass can be different. Sometimes a simple solution might actually be better. Luck is all we need.", "votes": 1 }, { "id": 3393172, "postDate": "2026-01-18T11:19:36.037Z", "content": "Even using semantic segmentation masks is not necessarily reliable, because the labeling method is unrelated to vision, while the model is only allowed to use vision to estimate. What the model learns may just be heuristics like: for this season and this location, the biomass value is usually around a certain level. Next year, if the climate conditions change or the location changes, the biomass values could be completely different.
", "rawMarkdown": "Even using semantic segmentation masks is not necessarily reliable, because the labeling method is unrelated to vision, while the model is only allowed to use vision to estimate. What the model learns may just be heuristics like: for this season and this location, the biomass value is usually around a certain level. Next year, if the climate conditions change or the location changes, the biomass values could be completely different.", "votes": 1 }, { "id": 3393177, "postDate": "2026-01-18T11:39:45.070Z", "content": "I am not too sure about a shake up in this competition, because the metric is quite hard to improve. Except if top solutions were overfitting on the public leaderboard
", "rawMarkdown": "I am not too sure about a shake up in this competition, because the metric is quite hard to improve. Except if top solutions were overfitting on the public leaderboard", "votes": 3 }, { "id": 3393180, "postDate": "2026-01-18T11:52:19.350Z", "content": "This is not related to our discussion, but it’s an example of betting on a more general solution rather than one that has been finely tuned. I can imagine what a 0.71 solution would look like, and in fact there’s no reason to dismiss a solution like that. Sometimes we have to accept the reality that a solution carefully tuned over three months can’t compare to one built in just a week.
", "rawMarkdown": "This is not related to our discussion, but it’s an [example](https://www.kaggle.com/competitions/csiro-biomass/discussion/666991#3392951) of betting on a more general solution rather than one that has been finely tuned. I can imagine what a 0.71 solution would look like, and in fact there’s no reason to dismiss a solution like that. Sometimes we have to accept the reality that a solution carefully tuned over three months can’t compare to one built in just a week.", "votes": 1 }, { "id": 3393710, "postDate": "2026-01-19T16:25:42.317Z", "content": "ya shake could happen cv and lb for even single models are all over the place .. a single model could win it also but i feel the shake will be between 74 -79 ranges where confidence is going to be low as the cv lb correlation is running so close
", "rawMarkdown": "ya shake could happen cv and lb for even single models are all over the place .. a single model could win it also but i feel the shake will be between 74 -79 ranges where confidence is going to be low as the cv lb correlation is running so close ", "votes": 2 } ] }, { "id": 3393233, "postDate": "2026-01-18T14:26:36.547Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3394507, "postDate": "2026-01-21T06:40:46.037Z", "content": "I agree @quan0095
", "rawMarkdown": "I agree @quan0095 " } ] }, { "id": 3393127, "postDate": "2026-01-18T09:39:45.520Z", "content": "I believe that the high-scoring public notebooks are overfitting to the leaderboard.\nSelecting specific folds and applying a fixed scaling factor as post-processing might still work on the private leaderboard, but I think it carries significant risk.\nhttps://www.kaggle.com/code/mayukh18/dinov3-no-tta-postprocess
", "rawMarkdown": "I believe that the high-scoring public notebooks are overfitting to the leaderboard.\nSelecting specific folds and applying a fixed scaling factor as post-processing might still work on the private leaderboard, but I think it carries significant risk.\nhttps://www.kaggle.com/code/mayukh18/dinov3-no-tta-postprocess", "replies": [ { "id": 3393128, "postDate": "2026-01-18T09:41:02.707Z", "rawMarkdown": "", "votes": -2, "isDeleted": true }, { "id": 3393135, "postDate": "2026-01-18T10:09:57.343Z", "content": "The same thing might happen as with the ICR competitions.
", "rawMarkdown": "The same thing might happen as with the ICR competitions.", "votes": -2 }, { "id": 3393497, "postDate": "2026-01-19T06:31:21.483Z", "content": "I think scaling factor makes no sense. Unless someone can magically gusse the true distribution in private testset.
", "rawMarkdown": "I think scaling factor makes no sense. Unless someone can magically gusse the true distribution in private testset.", "votes": 7, "replies": [ { "id": 3393738, "postDate": "2026-01-19T17:15:16.340Z", "content": "not scaling, not seeding, done without probing. am really waiting for the comp to end to see what the simple implementation was ;)
", "rawMarkdown": "not scaling, not seeding, done without probing. am really waiting for the comp to end to see what the simple implementation was ;)", "votes": 2 } ] }, { "id": 3393901, "postDate": "2026-01-20T05:23:24.527Z", "content": "There was previous discussion about samples containing clover, but Dry_Clover_g = 0. This might explain why Dry_Clover_g * 0.8 is effective. However, I still believe that scaling factor is a high-risk gamble.
", "rawMarkdown": "There was previous [discussion](https://www.kaggle.com/competitions/csiro-biomass/discussion/631335#3342435) about samples containing clover, but Dry_Clover_g = 0. This might explain why [Dry_Clover_g * 0.8](https://www.kaggle.com/code/mayukh18/dinov3-no-tta-postprocess) is effective. However, I still believe that scaling factor is a high-risk gamble.", "votes": 3, "replies": [ { "id": 3393946, "postDate": "2026-01-20T07:05:51.937Z", "content": "Let me try Dry_Clover_g * 0.8 🤗 I’m wondering whether this works with ensemble models, because with an ensemble the variance has already been reduced and it converges to the bias. If this has an effect, then it must be a bias in the data. Since we’re not sure about the root cause, we also can’t be sure whether data from other years will have the same type of bias.
", "rawMarkdown": "Let me try Dry_Clover_g * 0.8 🤗 I’m wondering whether this works with ensemble models, because with an ensemble the variance has already been reduced and it converges to the bias. If this has an effect, then it must be a bias in the data. Since we’re not sure about the root cause, we also can’t be sure whether data from other years will have the same type of bias.", "votes": 3 } ] } ] }, { "id": 3393869, "postDate": "2026-01-20T03:01:58.063Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3394757, "author_name": "agro_pixel_26", "author_url": "", "post_date": "2026-01-21T16:23:00.217000", "content": "\"I totally agree with you! The 'Late Submission Dumping' takes away the real learning spirit of the competition. It's frustrating to see hard-earned ranks slip due to public notebooks. Keep pushing, being in the top 250 by your own effort is much more valuable for your AGRO_TECH skills than just clicking 'run' on a public script! Upvoted for visibility.\"
", "votes": 4, "replies": [] }, { "id": 3393083, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-01-18T08:00:38.123000", "content": "Kaggle competitions are becoming increasingly difficult; in this competition, a shake-up is likely, and the private score of teams in the public gold zone could end up lower than the private score of a public notebook 🥱🥱🥱
", "votes": 2, "replies": [ { "id": 3393094, "author_name": "William Alabi", "author_url": "", "post_date": "2026-01-18T08:11:21.123000", "content": "How sure are you there would be a shake?
", "votes": 0, "replies": [ { "id": 3393163, "author_name": "Raj Raman83", "author_url": "", "post_date": "2026-01-18T11:00:52.443000", "content": "one reason could be using scaling factor i think, as the scaling factor would only for now , not when whole test set is used causing wrong total and etc
", "votes": 0, "replies": [] }, { "id": 3393164, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-01-18T11:03:32.387000", "content": "There is too little data, while the value ranges for dead, green, and clover are very wide. Meanwhile, the host split the data by year, and depending on the year, the value distribution of biomass can be different. Sometimes a simple solution might actually be better. Luck is all we need.
", "votes": 1, "replies": [] }, { "id": 3393172, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-01-18T11:19:36.037000", "content": "Even using semantic segmentation masks is not necessarily reliable, because the labeling method is unrelated to vision, while the model is only allowed to use vision to estimate. What the model learns may just be heuristics like: for this season and this location, the biomass value is usually around a certain level. Next year, if the climate conditions change or the location changes, the biomass values could be completely different.
", "votes": 1, "replies": [] }, { "id": 3393177, "author_name": "William Alabi", "author_url": "", "post_date": "2026-01-18T11:39:45.070000", "content": "I am not too sure about a shake up in this competition, because the metric is quite hard to improve. Except if top solutions were overfitting on the public leaderboard
", "votes": 3, "replies": [] }, { "id": 3393180, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-01-18T11:52:19.350000", "content": "This is not related to our discussion, but it’s an example of betting on a more general solution rather than one that has been finely tuned. I can imagine what a 0.71 solution would look like, and in fact there’s no reason to dismiss a solution like that. Sometimes we have to accept the reality that a solution carefully tuned over three months can’t compare to one built in just a week.
", "votes": 1, "replies": [] }, { "id": 3393710, "author_name": "Gaurav Rawat", "author_url": "", "post_date": "2026-01-19T16:25:42.317000", "content": "ya shake could happen cv and lb for even single models are all over the place .. a single model could win it also but i feel the shake will be between 74 -79 ranges where confidence is going to be low as the cv lb correlation is running so close
", "votes": 2, "replies": [] } ] }, { "id": 3393233, "author_name": "", "author_url": "", "post_date": "2026-01-18T14:26:36.547000", "content": "", "votes": 0, "replies": [] }, { "id": 3394507, "author_name": "Navneet", "author_url": "", "post_date": "2026-01-21T06:40:46.037000", "content": "I agree @quan0095
", "votes": 0, "replies": [] } ] }, { "id": 3393127, "author_name": "hiro", "author_url": "", "post_date": "2026-01-18T09:39:45.520000", "content": "I believe that the high-scoring public notebooks are overfitting to the leaderboard.\nSelecting specific folds and applying a fixed scaling factor as post-processing might still work on the private leaderboard, but I think it carries significant risk.\nhttps://www.kaggle.com/code/mayukh18/dinov3-no-tta-postprocess
", "votes": 0, "replies": [ { "id": 3393128, "author_name": "", "author_url": "", "post_date": "2026-01-18T09:41:02.707000", "content": "", "votes": -2, "replies": [] }, { "id": 3393135, "author_name": "Giovanny Rodríguez", "author_url": "", "post_date": "2026-01-18T10:09:57.343000", "content": "The same thing might happen as with the ICR competitions.
", "votes": -2, "replies": [] }, { "id": 3393497, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-19T06:31:21.483000", "content": "I think scaling factor makes no sense. Unless someone can magically gusse the true distribution in private testset.
", "votes": 7, "replies": [ { "id": 3393738, "author_name": "Taha_Alshatiri", "author_url": "", "post_date": "2026-01-19T17:15:16.340000", "content": "not scaling, not seeding, done without probing. am really waiting for the comp to end to see what the simple implementation was ;)
", "votes": 2, "replies": [] } ] }, { "id": 3393901, "author_name": "MOONMOON", "author_url": "", "post_date": "2026-01-20T05:23:24.527000", "content": "There was previous discussion about samples containing clover, but Dry_Clover_g = 0. This might explain why Dry_Clover_g * 0.8 is effective. However, I still believe that scaling factor is a high-risk gamble.
", "votes": 3, "replies": [ { "id": 3393946, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-01-20T07:05:51.937000", "content": "Let me try Dry_Clover_g * 0.8 🤗 I’m wondering whether this works with ensemble models, because with an ensemble the variance has already been reduced and it converges to the bias. If this has an effect, then it must be a bias in the data. Since we’re not sure about the root cause, we also can’t be sure whether data from other years will have the same type of bias.
", "votes": 3, "replies": [] } ] } ] }, { "id": 3393869, "author_name": "", "author_url": "", "post_date": "2026-01-20T03:01:58.063000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3393066": "Just a simple request .. I know i will get downvote for this but only few days are left in the competition and people are just posting the higher scoring submission notebook... It really effect the nature of the competition as many people are trying to score good by pushing hard on daily basis .. don't ruin this fun for everyone , it doesn't look good :)\n\nPS : Currently at 210th place and slipping on daily basis lol ", "3394757": "\"I totally agree with you! The 'Late Submission Dumping' takes away the real learning spirit of the competition. It's frustrating to see hard-earned ranks slip due to public notebooks. Keep pushing, being in the top 250 by your own effort is much more valuable for your AGRO_TECH skills than just clicking 'run' on a public script! Upvoted for visibility.\"", "3393083": "Kaggle competitions are becoming increasingly difficult; in this competition, a shake-up is likely, and the private score of teams in the public gold zone could end up lower than the private score of a public notebook 🥱🥱🥱", "3393127": "I believe that the high-scoring public notebooks are overfitting to the leaderboard.\nSelecting specific folds and applying a fixed scaling factor as post-processing might still work on the private leaderboard, but I think it carries significant risk.\nhttps://www.kaggle.com/code/mayukh18/dinov3-no-tta-postprocess", "3393869": "" } }