What is your best CV score? It seems to be easy to get 0.8+. My r2 scores are 0.63, 0.85, 0.87, 0.855, 0.79 for dead, clover, green, GDM and total. I think 0.84-0.85 is reachable. What do you think? \nI have not checked this solution on LB yet.
", "messages": [ { "id": "3389134", "postDate": "01/10/2026 11:23:33", "content": "What is your best CV score? It seems to be easy to get 0.8+. My r2 scores are 0.63, 0.85, 0.87, 0.855, 0.79 for dead, clover, green, GDM and total. I think 0.84-0.85 is reachable. What do you think? \nI have not checked this solution on LB yet.
", "rawMarkdown": "What is your best CV score? It seems to be easy to get 0.8+. My r2 scores are 0.63, 0.85, 0.87, 0.855, 0.79 for dead, clover, green, GDM and total. I think 0.84-0.85 is reachable. What do you think? \nI have not checked this solution on LB yet.", "votes": null }, { "id": "3389169", "postDate": "01/10/2026 12:17:34", "content": "I wish it's easy to reach 0.8 on LB.
", "rawMarkdown": "I wish it's easy to reach 0.8 on LB.", "votes": null }, { "id": "3389200", "postDate": "01/10/2026 13:56:15", "content": "My CV doesn't cross 0.73
", "rawMarkdown": "My CV doesn't cross 0.73", "votes": null }, { "id": "3389751", "postDate": "01/11/2026 20:18:17", "content": "I have reached .91 CV but I think there there was some leakage. Otherwise I get .80-.84 quite easily, but it doesn't really correlate to LB.
", "rawMarkdown": "I have reached .91 CV but I think there there was some leakage. Otherwise I get .80-.84 quite easily, but it doesn't really correlate to LB.", "votes": null }, { "id": "3389759", "postDate": "01/11/2026 20:36:42", "content": "I have 0.85 CV now, and it seems to be better and better with more training time. What kind of leakage it can be?
", "rawMarkdown": "I have 0.85 CV now, and it seems to be better and better with more training time. What kind of leakage it can be?", "votes": null }, { "id": "3389761", "postDate": "01/11/2026 20:42:24", "content": "For example there are images that are probably taken the same day in the same field. Some also have timestamps. There are discussions about the proper CV split method. The competition metric is also not exactly the average of the R2 per target.
", "rawMarkdown": "For example there are images that are probably taken the same day in the same field. Some also have timestamps. There are discussions about the proper CV split method. The competition metric is also not exactly the average of the R2 per target.", "votes": null }, { "id": "3389766", "postDate": "01/11/2026 20:50:59", "content": "I'm quite certain that there is multiple datapoints from the same farm/location, typically 4-6 very similar images with highly correlated targets. This is the main reason why people have better CV/LB correlation with GroupKFold (grouped by State&Date). Big risk for the model to just learn to recognize the spot and not the real targets.
\nEdit: tennogh was faster.
", "rawMarkdown": "I'm quite certain that there is multiple datapoints from the same farm/location, typically 4-6 very similar images with highly correlated targets. This is the main reason why people have better CV/LB correlation with GroupKFold (grouped by State&Date). Big risk for the model to just learn to recognize the spot and not the real targets.\n\nEdit: tennogh was faster.", "votes": null }, { "id": "3389775", "postDate": "01/11/2026 21:31:35", "content": "Yup, there are many images with this same state, sampling date and similar targets.
", "rawMarkdown": "Yup, there are many images with this same state, sampling date and similar targets.", "votes": null }, { "id": "3390245", "postDate": "01/12/2026 20:35:23", "content": "Do you use GroupKFold by date and state? It seems to be much harder than random kfolds.
", "rawMarkdown": "Do you use GroupKFold by date and state? It seems to be much harder than random kfolds.", "votes": null }, { "id": "3390344", "postDate": "01/13/2026 01:04:22", "content": "My CV is similar to yours, but the problem is that because the private leaderboard has a very limited number of samples, the CV results may not reliably represent the private leaderboard performance. Currently, I can only ensure that the trend in my CV is consistent with the public leaderboard scores.
", "rawMarkdown": "My CV is similar to yours, but the problem is that because the private leaderboard has a very limited number of samples, the CV results may not reliably represent the private leaderboard performance. Currently, I can only ensure that the trend in my CV is consistent with the public leaderboard scores.", "votes": null }, { "id": "3391651", "postDate": "01/15/2026 11:50:25", "content": "I use 5 fold training. when I use vit_base_dinov3, my local cv always between 0.83~0.84 and the LB between 0.66~0.69；when use vit_huge_plus_dinov3, my cv could touch 0.85, but huge's pth file more than 3G, so I have not upload it to see it's LB score. And I try to split the picture to 2patch and 8patch，I find 2patch's cv better than 8patch but 8patch's LB better than 2patch.
", "rawMarkdown": "I use 5 fold training. when I use vit_base_dinov3, my local cv always between 0.83~0.84 and the LB between 0.66~0.69；when use vit_huge_plus_dinov3, my cv could touch 0.85, but huge's pth file more than 3G, so I have not upload it to see it's LB score. And I try to split the picture to 2patch and 8patch，I find 2patch's cv better than 8patch but 8patch's LB better than 2patch.", "votes": null }, { "id": "3391803", "postDate": "01/15/2026 17:39:20", "content": "My CV scores are always 0.90+, but no single model can cross 0.65 on LB.
\nvery excited to see what sorcery the top ranked participants did to achieve 0.79 LB when they disclose their solutions 2 weeks later
", "rawMarkdown": "My CV scores are always 0.90+, but no single model can cross 0.65 on LB.\n\nvery excited to see what sorcery the top ranked participants did to achieve 0.79 LB when they disclose their solutions 2 weeks later", "votes": null }, { "id": "3392381", "postDate": "01/16/2026 20:35:21", "content": "Do you use huge k folds random splits?
", "rawMarkdown": "Do you use huge k folds random splits?", "votes": null }, { "id": "3392951", "postDate": "01/17/2026 22:16:43", "content": "Thanks for this small discussion. After one week of active kaggling, I have reached almost 0.71 LB score and this is enough. I don't see anything what could improve private LB. Respect your time and don't do 300+ submissions. Good luck!
", "rawMarkdown": "Thanks for this small discussion. After one week of active kaggling, I have reached almost 0.71 LB score and this is enough. I don't see anything what could improve private LB. Respect your time and don't do 300+ submissions. Good luck!", "votes": null }, { "id": "3392973", "postDate": "01/18/2026 00:50:34", "content": "you haven't tried this technique, implement this to see if it works for you: Mixup, NLL, Teacher (but for noise).
", "rawMarkdown": "you haven't tried this technique, implement this to see if it works for you: Mixup, NLL, Teacher (but for noise).", "votes": null } ], "comments": [ { "id": 3389169, "author_name": "medali1992", "author_url": "", "post_date": "01/10/2026 12:17:34", "content": "I wish it's easy to reach 0.8 on LB.
", "votes": null, "replies": [] }, { "id": 3389200, "author_name": "paritoshtripathi5", "author_url": "", "post_date": "01/10/2026 13:56:15", "content": "My CV doesn't cross 0.73
", "votes": null, "replies": [ { "id": 3390245, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/12/2026 20:35:23", "content": "Do you use GroupKFold by date and state? It seems to be much harder than random kfolds.
", "votes": null, "replies": [] } ] }, { "id": 3389751, "author_name": "tennogh", "author_url": "", "post_date": "01/11/2026 20:18:17", "content": "I have reached .91 CV but I think there there was some leakage. Otherwise I get .80-.84 quite easily, but it doesn't really correlate to LB.
", "votes": null, "replies": [ { "id": 3389759, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/11/2026 20:36:42", "content": "I have 0.85 CV now, and it seems to be better and better with more training time. What kind of leakage it can be?
", "votes": null, "replies": [ { "id": 3389761, "author_name": "tennogh", "author_url": "", "post_date": "01/11/2026 20:42:24", "content": "For example there are images that are probably taken the same day in the same field. Some also have timestamps. There are discussions about the proper CV split method. The competition metric is also not exactly the average of the R2 per target.
", "votes": null, "replies": [] }, { "id": 3389766, "author_name": "herrahuu", "author_url": "", "post_date": "01/11/2026 20:50:59", "content": "I'm quite certain that there is multiple datapoints from the same farm/location, typically 4-6 very similar images with highly correlated targets. This is the main reason why people have better CV/LB correlation with GroupKFold (grouped by State&Date). Big risk for the model to just learn to recognize the spot and not the real targets.
\nEdit: tennogh was faster.
", "votes": null, "replies": [ { "id": 3389775, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/11/2026 21:31:35", "content": "Yup, there are many images with this same state, sampling date and similar targets.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3390344, "author_name": "sweetyheehee", "author_url": "", "post_date": "01/13/2026 01:04:22", "content": "My CV is similar to yours, but the problem is that because the private leaderboard has a very limited number of samples, the CV results may not reliably represent the private leaderboard performance. Currently, I can only ensure that the trend in my CV is consistent with the public leaderboard scores.
", "votes": null, "replies": [] }, { "id": 3391651, "author_name": "zhiyue666", "author_url": "", "post_date": "01/15/2026 11:50:25", "content": "I use 5 fold training. when I use vit_base_dinov3, my local cv always between 0.83~0.84 and the LB between 0.66~0.69；when use vit_huge_plus_dinov3, my cv could touch 0.85, but huge's pth file more than 3G, so I have not upload it to see it's LB score. And I try to split the picture to 2patch and 8patch，I find 2patch's cv better than 8patch but 8patch's LB better than 2patch.
", "votes": null, "replies": [] }, { "id": 3391803, "author_name": "mushfirat", "author_url": "", "post_date": "01/15/2026 17:39:20", "content": "My CV scores are always 0.90+, but no single model can cross 0.65 on LB.
\nvery excited to see what sorcery the top ranked participants did to achieve 0.79 LB when they disclose their solutions 2 weeks later
", "votes": null, "replies": [ { "id": 3392381, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/16/2026 20:35:21", "content": "Do you use huge k folds random splits?
", "votes": null, "replies": [] } ] }, { "id": 3392951, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/17/2026 22:16:43", "content": "Thanks for this small discussion. After one week of active kaggling, I have reached almost 0.71 LB score and this is enough. I don't see anything what could improve private LB. Respect your time and don't do 300+ submissions. Good luck!
", "votes": null, "replies": [ { "id": 3392973, "author_name": "giovannyrodrguez", "author_url": "", "post_date": "01/18/2026 00:50:34", "content": "you haven't tried this technique, implement this to see if it works for you: Mixup, NLL, Teacher (but for noise).
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3389134": "What is your best CV score? It seems to be easy to get 0.8+. My r2 scores are 0.63, 0.85, 0.87, 0.855, 0.79 for dead, clover, green, GDM and total. I think 0.84-0.85 is reachable. What do you think? \nI have not checked this solution on LB yet.", "3389169": "I wish it's easy to reach 0.8 on LB.", "3389200": "My CV doesn't cross 0.73", "3389751": "I have reached .91 CV but I think there there was some leakage. Otherwise I get .80-.84 quite easily, but it doesn't really correlate to LB.", "3389759": "I have 0.85 CV now, and it seems to be better and better with more training time. What kind of leakage it can be?", "3389761": "For example there are images that are probably taken the same day in the same field. Some also have timestamps. There are discussions about the proper CV split method. The competition metric is also not exactly the average of the R2 per target.", "3389766": "I'm quite certain that there is multiple datapoints from the same farm/location, typically 4-6 very similar images with highly correlated targets. This is the main reason why people have better CV/LB correlation with GroupKFold (grouped by State&Date). Big risk for the model to just learn to recognize the spot and not the real targets.\n\nEdit: tennogh was faster.", "3389775": "Yup, there are many images with this same state, sampling date and similar targets.", "3390245": "Do you use GroupKFold by date and state? It seems to be much harder than random kfolds.", "3390344": "My CV is similar to yours, but the problem is that because the private leaderboard has a very limited number of samples, the CV results may not reliably represent the private leaderboard performance. Currently, I can only ensure that the trend in my CV is consistent with the public leaderboard scores.", "3391651": "I use 5 fold training. when I use vit_base_dinov3, my local cv always between 0.83~0.84 and the LB between 0.66~0.69；when use vit_huge_plus_dinov3, my cv could touch 0.85, but huge's pth file more than 3G, so I have not upload it to see it's LB score. And I try to split the picture to 2patch and 8patch，I find 2patch's cv better than 8patch but 8patch's LB better than 2patch.", "3391803": "My CV scores are always 0.90+, but no single model can cross 0.65 on LB.\n\nvery excited to see what sorcery the top ranked participants did to achieve 0.79 LB when they disclose their solutions 2 weeks later", "3392381": "Do you use huge k folds random splits?", "3392951": "Thanks for this small discussion. After one week of active kaggling, I have reached almost 0.71 LB score and this is enough. I don't see anything what could improve private LB. Respect your time and don't do 300+ submissions. Good luck!", "3392973": "you haven't tried this technique, implement this to see if it works for you: Mixup, NLL, Teacher (but for noise)." }, "source": "meta" }