![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15429833%2F4f001ee9282828dd64bbb35639515e4d%2F02d4e05b-e36c-43bb-b50e-1a7ca66f7511.png?generation=1769646249194515&alt=media\")
0.70 in public score more than 0.73 in public.
", "messages": [ { "id": 3398291, "postDate": "2026-01-29T00:24:49.250Z", "content": "
Just see what I've choose lol
\nFun fact: at the last day, I retrained some models on ViT7B+ and get 0.78 on public LB, then I submit them. If I did'n do this, I will submit another ensemble model with 0.67 private LB.
", "rawMarkdown": "\n\nJust see what I've choose lol\n\n____\n\nFun fact: at the last day, I retrained some models on ViT7B+ and get 0.78 on public LB, then I submit them. If I did'n do this, I will submit another ensemble model with 0.67 private LB.", "votes": 20, "replies": [ { "id": 3398293, "postDate": "2026-01-29T00:26:03.783Z", "content": "no way im so sorry lol
", "rawMarkdown": "no way im so sorry lol", "replies": [ { "id": 3398294, "postDate": "2026-01-29T00:26:17.697Z", "content": "thats so crazy
", "rawMarkdown": "thats so crazy" } ] }, { "id": 3398295, "postDate": "2026-01-29T00:27:15.307Z", "content": "Thanks… I feel better now lol
", "rawMarkdown": "Thanks... I feel better now lol" }, { "id": 3398296, "postDate": "2026-01-29T00:27:49.357Z", "content": "coldest comment
", "rawMarkdown": "coldest comment" }, { "id": 3398297, "postDate": "2026-01-29T00:29:37.203Z", "content": "Haha. Was the 0.69 solution the one you said was a very simple solution?
", "rawMarkdown": "Haha. Was the 0.69 solution the one you said was a very simple solution?", "replies": [ { "id": 3398300, "postDate": "2026-01-29T00:30:45.767Z", "content": "Yes, indeed.
", "rawMarkdown": "Yes, indeed." } ] }, { "id": 3398302, "postDate": "2026-01-29T00:31:11.517Z", "content": "You really cracked it 👏 . interesting to know your folds and solution :) .. Unfortunately for us though cv lb corr was ok the scores didnt work at all in private :D . Maybe hence I had named our team frustrating folds
", "rawMarkdown": "You really cracked it 👏 . interesting to know your folds and solution :) .. Unfortunately for us though cv lb corr was ok the scores didnt work at all in private :D . Maybe hence I had named our team frustrating folds", "replies": [ { "id": 3398306, "postDate": "2026-01-29T00:35:22.517Z", "content": "For me, CV fits the final private LB well. My best CV is about 0.67.
", "rawMarkdown": "For me, CV fits the final private LB well. My best CV is about 0.67." }, { "id": 3398308, "postDate": "2026-01-29T00:37:06.720Z", "content": "yes you got the cv strategy right .. Though lot of learnign will try to see how our model perform in those cv folds . as feel our archs were good but will learn soon . Thanks .
", "rawMarkdown": "yes you got the cv strategy right .. Though lot of learnign will try to see how our model perform in those cv folds . as feel our archs were good but will learn soon . Thanks . \n" }, { "id": 3398426, "postDate": "2026-01-29T05:54:34.560Z", "content": "\n\nFor me, CV fits the final private LB well. My best CV is about 0.67.
\n
Really looking for your writeup!
", "rawMarkdown": "> For me, CV fits the final private LB well. My best CV is about 0.67.\n\n\nReally looking for your writeup!", "votes": 1 } ] }, { "id": 3398305, "postDate": "2026-01-29T00:35:04.147Z", "content": "Wow, amazing. Please give insights about these submissions also in your solution writeup. It will help everyone a lot.
", "rawMarkdown": "Wow, amazing. Please give insights about these submissions also in your solution writeup. It will help everyone a lot." }, { "id": 3398313, "postDate": "2026-01-29T00:50:42.837Z", "content": "I have a genuine question as a student: how should I interpret my results, or in other words, how should I feel about them?
\nMy best submission (which I did not select) had several higher scores 0.63, but none reached 0.64. I also experienced some of the smallest shake-ups between public and private. In fact, I also mentioned that I am not expecting a shakeup at least for myself.
\nOn one hand, I ended up surrounded by the public leaderboard toppers:
\nOn the other hand, I see people who were ranked 100+ or even 800+ on public who scored higher than me on private. I’m curious: did some of them intentionally stay lower on the public LB (for example around 100+) expecting this kind of shake-up, or was it more a matter of luck versus good execution?
\nFor context, I mentioned several times that my training R2 and LB scores correlated almost perfectly only up to about 0.72 R2. The checkpoint I finally selected is exactly that one, and it outperforms the 0.74–0.75 R2 checkpoints, which were clearly overfitting.
\nI’m not complaining at all, I’m just trying to learn from this experience. Given all this, do you think this is something I should feel proud of and build on?
", "rawMarkdown": "I have a genuine question as a student: how should I interpret my results, or in other words, how should I feel about them? \n\nMy best submission (which I did not select) had several higher scores 0.63, but none reached 0.64. I also experienced some of the smallest shake-ups between public and private. In fact, I also mentioned that **I am not expecting a shakeup at least for myself.**\n\nOn one hand, I ended up surrounded by the public leaderboard toppers: \n- 2 → 39 \n- 3 → 40 \n- 49 → 41 (Me)\n- 1 → 42 \n\nOn the other hand, I see people who were ranked 100+ or even 800+ on public who scored higher than me on private. I’m curious: did some of them intentionally stay lower on the public LB (for example around 100+) expecting this kind of shake-up, or was it more a matter of luck versus good execution? \n\nFor context, I mentioned several times that my training R2 and LB scores correlated almost perfectly only up to about 0.72 R2. The checkpoint I finally selected is exactly that one, and it outperforms the 0.74–0.75 R2 checkpoints, which were clearly overfitting. \n\nI’m not complaining at all, I’m just trying to learn from this experience. Given all this, do you think this is something I should feel proud of and build on?", "votes": 1, "replies": [ { "id": 3398320, "postDate": "2026-01-29T01:05:23.693Z", "content": "I think anyone committed on this contest should be proud of him/herself. True luck will affect your rank, but rank is not the meaning of this contest. I've learn a lot idea on this contest(and apply some of them in my work then get well results), that's enough.
", "rawMarkdown": "I think anyone committed on this contest should be proud of him/herself. True luck will affect your rank, but rank is not the meaning of this contest. I've learn a lot idea on this contest(and apply some of them in my work then get well results), that's enough.", "votes": 7 }, { "id": 3398345, "postDate": "2026-01-29T02:08:30.847Z", "content": "I'm one of +821 ranking. To be honest, I’m not surprised by the result. A Silver medal on the Private LB still feels very achievable even with a 1000+ rank on the Public LB. In contrast, the gold medalists will likely be concentrated within roughly the top ±50 positions, as they usually have very stable performance across CV, Public LB
\nI based my decision and hyperparams tunning on 16-folds CV (4-folds splitting with 4 different scheme), 8x TTA for each fold, result in totally 32x size OOF metrics and just submit best CV model to LB.
\nMore things \"that should work in theory\" but did not transfer to LB improvement. It's just too risky for me for a gold to put more time in, so I stop competing 24 days ago.
\nMatter of luck: yes, huge factor for non-top50 LB like me :D
", "rawMarkdown": "I'm one of +821 ranking. To be honest, I’m not surprised by the result. A Silver medal on the Private LB still feels very achievable even with a 1000+ rank on the Public LB. In contrast, the gold medalists will likely be concentrated within roughly the top ±50 positions, as they usually have very stable performance across CV, Public LB\n\nI based my decision and hyperparams tunning on 16-folds CV (4-folds splitting with 4 different scheme), 8x TTA for each fold, result in totally 32x size OOF metrics and just submit best CV model to LB.\n\nMore things \"that should work in theory\" but did not transfer to LB improvement. It's just too risky for me for a gold to put more time in, so I stop competing 24 days ago.\n\nMatter of luck: yes, huge factor for non-top50 LB like me :D", "votes": 5 } ] }, { "id": 3398322, "postDate": "2026-01-29T01:17:53.163Z", "content": "so I guess the simple way to get hige score is train a large yet well-suited model (like ViT7B+) and large image size?
", "rawMarkdown": "so I guess the simple way to get hige score is train a large yet well-suited model (like ViT7B+) and large image size?", "replies": [ { "id": 3398325, "postDate": "2026-01-29T01:26:44.713Z", "content": "vith16+ is enough. Train a simple head with strong backbone like dinov3, and do PEFT on it (like LoRA).
", "rawMarkdown": "vith16+ is enough. Train a simple head with strong backbone like dinov3, and do PEFT on it (like LoRA).", "votes": 2 }, { "id": 3398327, "postDate": "2026-01-29T01:32:55.243Z", "content": "i never got good scores from using dinov3 but good scores from dinov2. i wonder how you use dinov3
", "rawMarkdown": "i never got good scores from using dinov3 but good scores from dinov2. i wonder how you use dinov3", "votes": 1 }, { "id": 3398328, "postDate": "2026-01-29T01:37:30.080Z", "content": "We used ViT-7B. Not true.
", "rawMarkdown": "We used ViT-7B. Not true." }, { "id": 3398333, "postDate": "2026-01-29T01:42:35.927Z", "content": "I never touched the backbone and used only the Large variant. I used VPT (Visual Prompt Tuning).
", "rawMarkdown": "I never touched the backbone and used only the Large variant. I used VPT (Visual Prompt Tuning)." }, { "id": 3398336, "postDate": "2026-01-29T01:56:42.787Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3398337, "postDate": "2026-01-29T01:57:23.633Z", "content": "I use vit huge-dinov3 full fine-tuning can get 0.75LB and 0.63PB. I split image into 8 patch/tiles as input.
", "rawMarkdown": "I use vit huge-dinov3 full fine-tuning can get 0.75LB and 0.63PB. I split image into 8 patch/tiles as input." }, { "id": 3398339, "postDate": "2026-01-29T02:01:55.923Z", "content": "ya vit dinov3 huge I feel was enough.. ofc our thing dint work in private :) .. but yes it I feel was enough
", "rawMarkdown": "ya vit dinov3 huge I feel was enough.. ofc our thing dint work in private :) .. but yes it I feel was enough" }, { "id": 3398355, "postDate": "2026-01-29T02:46:59.580Z", "content": "I've publish my code: https://www.kaggle.com/code/linkinpony/dinov3-regressor. But to be honest, they are totally generated by AI.
", "rawMarkdown": "I've publish my code: https://www.kaggle.com/code/linkinpony/dinov3-regressor. But to be honest, they are totally generated by AI.", "votes": 3 }, { "id": 3398371, "postDate": "2026-01-29T03:28:25.947Z", "content": "How long did it take to train your models? my dinov2 took 9hours using kaggle notebook with lr = 5e-3.
", "rawMarkdown": "How long did it take to train your models? my dinov2 took 9hours using kaggle notebook with lr = 5e-3." }, { "id": 3398376, "postDate": "2026-01-29T03:33:12.983Z", "content": "20-90 mins on 4090
", "rawMarkdown": "20-90 mins on 4090" } ] }, { "id": 3398525, "postDate": "2026-01-29T09:11:45.420Z", "content": "OMG! that scores is crazy, so sorry
\nlooking forward to reading your overall approach.
\nFor me I figured out very late (just 3 days ago) the fact that large pretrainted superior features what worked here. \nI was trying to finetune smaller ConvNext-atto model with heads and guided attention from segmentation maps and feature engineerd blocks lol ><
\ngood that I figured out this at last moment ><
", "rawMarkdown": "OMG! that scores is crazy, so sorry\n\nlooking forward to reading your overall approach.\n\nFor me I figured out very late (just 3 days ago) the fact that large pretrainted superior features what worked here. \nI was trying to finetune smaller ConvNext-atto model with heads and guided attention from segmentation maps and feature engineerd blocks lol ><\n\ngood that I figured out this at last moment ><\n" }, { "id": 3398818, "postDate": "2026-01-29T18:34:08.657Z", "content": "Bro, We are sorry for your loss. But you stand strong you gonna win more of these, send a team invite for some competition 😅
", "rawMarkdown": "Bro, We are sorry for your loss. But you stand strong you gonna win more of these, send a team invite for some competition 😅" } ] }, { "id": 3398286, "postDate": "2026-01-29T00:09:53.347Z", "content": "0.70 in public score more than 0.73 in public.
", "rawMarkdown": "0.70 in public score more than 0.73 in public.", "votes": 5 }, { "id": 3398312, "postDate": "2026-01-29T00:47:49.943Z", "content": "🥲 I had a 0.63 in there
", "rawMarkdown": "🥲 I had a 0.63 in there ", "votes": 3 }, { "id": 3398352, "postDate": "2026-01-29T02:41:28.753Z", "content": "Come on, in actual production scenarios, you often don't even know what kind of dataset you're dealing with, and selecting the appropriate model is a reflection of individual expertise.
", "rawMarkdown": "Come on, in actual production scenarios, you often don't even know what kind of dataset you're dealing with, and selecting the appropriate model is a reflection of individual expertise.", "votes": 2 }, { "id": 3398914, "postDate": "2026-01-29T22:26:15.320Z", "content": "I am curious, why did the top 3 on the public leaderboard all shake down by the same magnitude? Is there only one way to get a 0.79+ on public and they have the same degree of overfitting?
", "rawMarkdown": "I am curious, why did the top 3 on the public leaderboard all shake down by the **same magnitude**? Is there only one way to get a 0.79+ on public and they have the same degree of overfitting?" }, { "id": 3398340, "postDate": "2026-01-29T02:02:05.013Z", "content": "me two, this is so bad😭😪
", "rawMarkdown": "me two, this is so bad😭😪" }, { "id": 3398319, "postDate": "2026-01-29T00:59:14.167Z", "content": "Wish the auto select, used the private score and not the public. Still wouldn't have won but would've moved up a few. \nCongratulations to the winner and condolences to those who selected a lower performing sub.
", "rawMarkdown": "Wish the auto select, used the private score and not the public. Still wouldn't have won but would've moved up a few. \nCongratulations to the winner and condolences to those who selected a lower performing sub.", "replies": [ { "id": 3398522, "postDate": "2026-01-29T09:05:53.343Z", "content": "I think the reason for 2 slection is that scoring 100k submissions would take too much infrastrucutre. So auto select private score they need to burn lot of compute which not sensible at that point there no point to even selecting submission
", "rawMarkdown": "I think the reason for 2 slection is that scoring 100k submissions would take too much infrastrucutre. So auto select private score they need to burn lot of compute which not sensible at that point there no point to even selecting submission", "replies": [ { "id": 3398671, "postDate": "2026-01-29T14:12:40.330Z", "content": "they do all the computation simultaneously, kaggle stuff can see all the scores before the competition end. If one could see the private score they can simply overfit towards it
", "rawMarkdown": "they do all the computation simultaneously, kaggle stuff can see all the scores before the competition end. If one could see the private score they can simply overfit towards it ", "votes": 2 }, { "id": 3398676, "postDate": "2026-01-29T14:20:34.457Z", "content": "ohh! I thought it gets computed later. Thank you for clarifying
", "rawMarkdown": "ohh! I thought it gets computed later. Thank you for clarifying " } ] } ] }, { "id": 3398290, "postDate": "2026-01-29T00:19:36.590Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3398291, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T00:24:49.250000", "content": "
Just see what I've choose lol
\nFun fact: at the last day, I retrained some models on ViT7B+ and get 0.78 on public LB, then I submit them. If I did'n do this, I will submit another ensemble model with 0.67 private LB.
", "votes": 20, "replies": [ { "id": 3398293, "author_name": "evnsnclr", "author_url": "", "post_date": "2026-01-29T00:26:03.783000", "content": "no way im so sorry lol
", "votes": 0, "replies": [ { "id": 3398294, "author_name": "evnsnclr", "author_url": "", "post_date": "2026-01-29T00:26:17.697000", "content": "thats so crazy
", "votes": 0, "replies": [] } ] }, { "id": 3398295, "author_name": "Yannan Chen", "author_url": "", "post_date": "2026-01-29T00:27:15.307000", "content": "Thanks… I feel better now lol
", "votes": 0, "replies": [] }, { "id": 3398296, "author_name": "zera", "author_url": "", "post_date": "2026-01-29T00:27:49.357000", "content": "coldest comment
", "votes": 0, "replies": [] }, { "id": 3398297, "author_name": "Mayukh Bhattacharyya", "author_url": "", "post_date": "2026-01-29T00:29:37.203000", "content": "Haha. Was the 0.69 solution the one you said was a very simple solution?
", "votes": 0, "replies": [ { "id": 3398300, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T00:30:45.767000", "content": "Yes, indeed.
", "votes": 0, "replies": [] } ] }, { "id": 3398302, "author_name": "Gaurav Rawat", "author_url": "", "post_date": "2026-01-29T00:31:11.517000", "content": "You really cracked it 👏 . interesting to know your folds and solution :) .. Unfortunately for us though cv lb corr was ok the scores didnt work at all in private :D . Maybe hence I had named our team frustrating folds
", "votes": 0, "replies": [ { "id": 3398306, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T00:35:22.517000", "content": "For me, CV fits the final private LB well. My best CV is about 0.67.
", "votes": 0, "replies": [] }, { "id": 3398308, "author_name": "Gaurav Rawat", "author_url": "", "post_date": "2026-01-29T00:37:06.720000", "content": "yes you got the cv strategy right .. Though lot of learnign will try to see how our model perform in those cv folds . as feel our archs were good but will learn soon . Thanks .
", "votes": 0, "replies": [] }, { "id": 3398426, "author_name": "Đăng Nguyễn Hồng", "author_url": "", "post_date": "2026-01-29T05:54:34.560000", "content": "\n\nFor me, CV fits the final private LB well. My best CV is about 0.67.
\n
Really looking for your writeup!
", "votes": 1, "replies": [] } ] }, { "id": 3398305, "author_name": "Komil Parmar", "author_url": "", "post_date": "2026-01-29T00:35:04.147000", "content": "Wow, amazing. Please give insights about these submissions also in your solution writeup. It will help everyone a lot.
", "votes": 0, "replies": [] }, { "id": 3398313, "author_name": "Komil Parmar", "author_url": "", "post_date": "2026-01-29T00:50:42.837000", "content": "I have a genuine question as a student: how should I interpret my results, or in other words, how should I feel about them?
\nMy best submission (which I did not select) had several higher scores 0.63, but none reached 0.64. I also experienced some of the smallest shake-ups between public and private. In fact, I also mentioned that I am not expecting a shakeup at least for myself.
\nOn one hand, I ended up surrounded by the public leaderboard toppers:
\nOn the other hand, I see people who were ranked 100+ or even 800+ on public who scored higher than me on private. I’m curious: did some of them intentionally stay lower on the public LB (for example around 100+) expecting this kind of shake-up, or was it more a matter of luck versus good execution?
\nFor context, I mentioned several times that my training R2 and LB scores correlated almost perfectly only up to about 0.72 R2. The checkpoint I finally selected is exactly that one, and it outperforms the 0.74–0.75 R2 checkpoints, which were clearly overfitting.
\nI’m not complaining at all, I’m just trying to learn from this experience. Given all this, do you think this is something I should feel proud of and build on?
", "votes": 1, "replies": [ { "id": 3398320, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T01:05:23.693000", "content": "I think anyone committed on this contest should be proud of him/herself. True luck will affect your rank, but rank is not the meaning of this contest. I've learn a lot idea on this contest(and apply some of them in my work then get well results), that's enough.
", "votes": 7, "replies": [] }, { "id": 3398345, "author_name": "Đăng Nguyễn Hồng", "author_url": "", "post_date": "2026-01-29T02:08:30.847000", "content": "I'm one of +821 ranking. To be honest, I’m not surprised by the result. A Silver medal on the Private LB still feels very achievable even with a 1000+ rank on the Public LB. In contrast, the gold medalists will likely be concentrated within roughly the top ±50 positions, as they usually have very stable performance across CV, Public LB
\nI based my decision and hyperparams tunning on 16-folds CV (4-folds splitting with 4 different scheme), 8x TTA for each fold, result in totally 32x size OOF metrics and just submit best CV model to LB.
\nMore things \"that should work in theory\" but did not transfer to LB improvement. It's just too risky for me for a gold to put more time in, so I stop competing 24 days ago.
\nMatter of luck: yes, huge factor for non-top50 LB like me :D
", "votes": 5, "replies": [] } ] }, { "id": 3398322, "author_name": "!!!!!", "author_url": "", "post_date": "2026-01-29T01:17:53.163000", "content": "so I guess the simple way to get hige score is train a large yet well-suited model (like ViT7B+) and large image size?
", "votes": 0, "replies": [ { "id": 3398325, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T01:26:44.713000", "content": "vith16+ is enough. Train a simple head with strong backbone like dinov3, and do PEFT on it (like LoRA).
", "votes": 2, "replies": [] }, { "id": 3398327, "author_name": "MengYe", "author_url": "", "post_date": "2026-01-29T01:32:55.243000", "content": "i never got good scores from using dinov3 but good scores from dinov2. i wonder how you use dinov3
", "votes": 1, "replies": [] }, { "id": 3398328, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-01-29T01:37:30.080000", "content": "We used ViT-7B. Not true.
", "votes": 0, "replies": [] }, { "id": 3398333, "author_name": "Komil Parmar", "author_url": "", "post_date": "2026-01-29T01:42:35.927000", "content": "I never touched the backbone and used only the Large variant. I used VPT (Visual Prompt Tuning).
", "votes": 0, "replies": [] }, { "id": 3398336, "author_name": "", "author_url": "", "post_date": "2026-01-29T01:56:42.787000", "content": "", "votes": 0, "replies": [] }, { "id": 3398337, "author_name": "!!!!!", "author_url": "", "post_date": "2026-01-29T01:57:23.633000", "content": "I use vit huge-dinov3 full fine-tuning can get 0.75LB and 0.63PB. I split image into 8 patch/tiles as input.
", "votes": 0, "replies": [] }, { "id": 3398339, "author_name": "Gaurav Rawat", "author_url": "", "post_date": "2026-01-29T02:01:55.923000", "content": "ya vit dinov3 huge I feel was enough.. ofc our thing dint work in private :) .. but yes it I feel was enough
", "votes": 0, "replies": [] }, { "id": 3398355, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T02:46:59.580000", "content": "I've publish my code: https://www.kaggle.com/code/linkinpony/dinov3-regressor. But to be honest, they are totally generated by AI.
", "votes": 3, "replies": [] }, { "id": 3398371, "author_name": "MengYe", "author_url": "", "post_date": "2026-01-29T03:28:25.947000", "content": "How long did it take to train your models? my dinov2 took 9hours using kaggle notebook with lr = 5e-3.
", "votes": 0, "replies": [] }, { "id": 3398376, "author_name": "Sayaka Miki", "author_url": "", "post_date": "2026-01-29T03:33:12.983000", "content": "20-90 mins on 4090
", "votes": 0, "replies": [] } ] }, { "id": 3398525, "author_name": "Pushpak Bhoge", "author_url": "", "post_date": "2026-01-29T09:11:45.420000", "content": "OMG! that scores is crazy, so sorry
\nlooking forward to reading your overall approach.
\nFor me I figured out very late (just 3 days ago) the fact that large pretrainted superior features what worked here. \nI was trying to finetune smaller ConvNext-atto model with heads and guided attention from segmentation maps and feature engineerd blocks lol ><
\ngood that I figured out this at last moment ><
", "votes": 0, "replies": [] }, { "id": 3398818, "author_name": "Bhargavaram Gajula", "author_url": "", "post_date": "2026-01-29T18:34:08.657000", "content": "Bro, We are sorry for your loss. But you stand strong you gonna win more of these, send a team invite for some competition 😅
", "votes": 0, "replies": [] } ] }, { "id": 3398312, "author_name": "Jaison Saji Chacko", "author_url": "", "post_date": "2026-01-29T00:47:49.943000", "content": "🥲 I had a 0.63 in there
", "votes": 3, "replies": [] }, { "id": 3398352, "author_name": "Timmy Juicehouse", "author_url": "", "post_date": "2026-01-29T02:41:28.753000", "content": "Come on, in actual production scenarios, you often don't even know what kind of dataset you're dealing with, and selecting the appropriate model is a reflection of individual expertise.
", "votes": 2, "replies": [] }, { "id": 3398914, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-01-29T22:26:15.320000", "content": "I am curious, why did the top 3 on the public leaderboard all shake down by the same magnitude? Is there only one way to get a 0.79+ on public and they have the same degree of overfitting?
", "votes": 0, "replies": [] }, { "id": 3398340, "author_name": "Robert0921", "author_url": "", "post_date": "2026-01-29T02:02:05.013000", "content": "me two, this is so bad😭😪
", "votes": 0, "replies": [] }, { "id": 3398319, "author_name": "Adam Lloyd", "author_url": "", "post_date": "2026-01-29T00:59:14.167000", "content": "Wish the auto select, used the private score and not the public. Still wouldn't have won but would've moved up a few. \nCongratulations to the winner and condolences to those who selected a lower performing sub.
", "votes": 0, "replies": [ { "id": 3398522, "author_name": "Pushpak Bhoge", "author_url": "", "post_date": "2026-01-29T09:05:53.343000", "content": "I think the reason for 2 slection is that scoring 100k submissions would take too much infrastrucutre. So auto select private score they need to burn lot of compute which not sensible at that point there no point to even selecting submission
", "votes": 0, "replies": [ { "id": 3398671, "author_name": "Andrey Chankin", "author_url": "", "post_date": "2026-01-29T14:12:40.330000", "content": "they do all the computation simultaneously, kaggle stuff can see all the scores before the competition end. If one could see the private score they can simply overfit towards it
", "votes": 2, "replies": [] }, { "id": 3398676, "author_name": "Pushpak Bhoge", "author_url": "", "post_date": "2026-01-29T14:20:34.457000", "content": "ohh! I thought it gets computed later. Thank you for clarifying
", "votes": 0, "replies": [] } ] } ] }, { "id": 3398290, "author_name": "", "author_url": "", "post_date": "2026-01-29T00:19:36.590000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3398291": "\n\nJust see what I've choose lol\n\n____\n\nFun fact: at the last day, I retrained some models on ViT7B+ and get 0.78 on public LB, then I submit them. If I did'n do this, I will submit another ensemble model with 0.67 private LB.", "3398286": "0.70 in public score more than 0.73 in public.", "3398312": "🥲 I had a 0.63 in there ", "3398352": "Come on, in actual production scenarios, you often don't even know what kind of dataset you're dealing with, and selecting the appropriate model is a reflection of individual expertise.", "3398914": "I am curious, why did the top 3 on the public leaderboard all shake down by the **same magnitude**? Is there only one way to get a 0.79+ on public and they have the same degree of overfitting?", "3398340": "me two, this is so bad😭😪", "3398319": "Wish the auto select, used the private score and not the public. Still wouldn't have won but would've moved up a few. \nCongratulations to the winner and condolences to those who selected a lower performing sub.", "3398290": "" } }