Does this competition lean more towards professional data augmentation strategies or model structures?
", "messages": [ { "id": "3392499", "postDate": "01/17/2026 04:14:35", "content": "Does this competition lean more towards professional data augmentation strategies or model structures?
", "rawMarkdown": "Does this competition lean more towards professional data augmentation strategies or model structures?", "votes": null }, { "id": "3392508", "postDate": "01/17/2026 04:40:05", "content": "Both…..(Another thing is that the winner of this competition will likely use a tree-based model. From what I can see, the current models have a lot of overfitting..)
", "rawMarkdown": "Both.....(Another thing is that the winner of this competition will likely use a tree-based model. From what I can see, the current models have a lot of overfitting..)", "votes": null }, { "id": "3392522", "postDate": "01/17/2026 05:30:57", "content": "I think it’s both. I improved the model structure as well as the data augmentation.
", "rawMarkdown": "I think it’s both. I improved the model structure as well as the data augmentation.", "votes": null }, { "id": "3392578", "postDate": "01/17/2026 06:46:27", "content": "I am a beginner in CV. I have tried changing the model architecture and increasing or decreasing the use of some augmentations, but improvements in CV do not correspond to improvements in LB. So, do you think in this competition we should trust CV improvements or LB improvements? Even though both LB and PB use 50% of the data, I would be really grateful if you could give some advice.
", "rawMarkdown": "I am a beginner in CV. I have tried changing the model architecture and increasing or decreasing the use of some augmentations, but improvements in CV do not correspond to improvements in LB. So, do you think in this competition we should trust CV improvements or LB improvements? Even though both LB and PB use 50% of the data, I would be really grateful if you could give some advice.", "votes": null }, { "id": "3392590", "postDate": "01/17/2026 07:17:41", "content": "In my experiments, tree based models couldn't even come close let alone perform better.
", "rawMarkdown": "In my experiments, tree based models couldn't even come close let alone perform better.", "votes": null }, { "id": "3392605", "postDate": "01/17/2026 07:56:29", "content": "I'm only saying this because of the overfitting, yes, since the maximum I achieved in the tree model was 69 (which is much worse).
", "rawMarkdown": "I'm only saying this because of the overfitting, yes, since the maximum I achieved in the tree model was 69 (which is much worse).", "votes": null }, { "id": "3392700", "postDate": "01/17/2026 11:43:43", "content": "Also, apart from the two mentioned paths, I believe the Training dynamics also play a significant role here. How much do you train, how do you avoid oscillations, when do you early stop, other optimizer and scheduling configurations, etc.
", "rawMarkdown": "Also, apart from the two mentioned paths, I believe the Training dynamics also play a significant role here. How much do you train, how do you avoid oscillations, when do you early stop, other optimizer and scheduling configurations, etc.", "votes": null }, { "id": "3393076", "postDate": "01/18/2026 07:49:34", "content": "In this task, we can’t rely solely on CV or the public leaderboard, because the distributions of the train, public, and private datasets are different. A solution is more reliable if it improves both CV and the public leaderboard, and if it is a reasonable, trustworthy approach. In addition, a reliable solution should have low variance across folds.
", "rawMarkdown": "In this task, we can’t rely solely on CV or the public leaderboard, because the distributions of the train, public, and private datasets are different. A solution is more reliable if it improves both CV and the public leaderboard, and if it is a reasonable, trustworthy approach. In addition, a reliable solution should have low variance across folds.", "votes": null }, { "id": "3393133", "postDate": "01/18/2026 09:58:50", "content": "now max: 71
", "rawMarkdown": "now max: 71", "votes": null } ], "comments": [ { "id": 3392508, "author_name": "giovannyrodrguez", "author_url": "", "post_date": "01/17/2026 04:40:05", "content": "Both…..(Another thing is that the winner of this competition will likely use a tree-based model. From what I can see, the current models have a lot of overfitting..)
", "votes": null, "replies": [ { "id": 3392590, "author_name": "paritoshtripathi5", "author_url": "", "post_date": "01/17/2026 07:17:41", "content": "In my experiments, tree based models couldn't even come close let alone perform better.
", "votes": null, "replies": [ { "id": 3392605, "author_name": "giovannyrodrguez", "author_url": "", "post_date": "01/17/2026 07:56:29", "content": "I'm only saying this because of the overfitting, yes, since the maximum I achieved in the tree model was 69 (which is much worse).
", "votes": null, "replies": [ { "id": 3393133, "author_name": "giovannyrodrguez", "author_url": "", "post_date": "01/18/2026 09:58:50", "content": "now max: 71
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3392522, "author_name": "quan0095", "author_url": "", "post_date": "01/17/2026 05:30:57", "content": "I think it’s both. I improved the model structure as well as the data augmentation.
", "votes": null, "replies": [ { "id": 3392578, "author_name": "muran169633", "author_url": "", "post_date": "01/17/2026 06:46:27", "content": "I am a beginner in CV. I have tried changing the model architecture and increasing or decreasing the use of some augmentations, but improvements in CV do not correspond to improvements in LB. So, do you think in this competition we should trust CV improvements or LB improvements? Even though both LB and PB use 50% of the data, I would be really grateful if you could give some advice.
", "votes": null, "replies": [ { "id": 3393076, "author_name": "quan0095", "author_url": "", "post_date": "01/18/2026 07:49:34", "content": "In this task, we can’t rely solely on CV or the public leaderboard, because the distributions of the train, public, and private datasets are different. A solution is more reliable if it improves both CV and the public leaderboard, and if it is a reasonable, trustworthy approach. In addition, a reliable solution should have low variance across folds.
", "votes": null, "replies": [] } ] }, { "id": 3392700, "author_name": "komilparmar", "author_url": "", "post_date": "01/17/2026 11:43:43", "content": "Also, apart from the two mentioned paths, I believe the Training dynamics also play a significant role here. How much do you train, how do you avoid oscillations, when do you early stop, other optimizer and scheduling configurations, etc.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3392499": "Does this competition lean more towards professional data augmentation strategies or model structures?", "3392508": "Both.....(Another thing is that the winner of this competition will likely use a tree-based model. From what I can see, the current models have a lot of overfitting..)", "3392522": "I think it’s both. I improved the model structure as well as the data augmentation.", "3392578": "I am a beginner in CV. I have tried changing the model architecture and increasing or decreasing the use of some augmentations, but improvements in CV do not correspond to improvements in LB. So, do you think in this competition we should trust CV improvements or LB improvements? Even though both LB and PB use 50% of the data, I would be really grateful if you could give some advice.", "3392590": "In my experiments, tree based models couldn't even come close let alone perform better.", "3392605": "I'm only saying this because of the overfitting, yes, since the maximum I achieved in the tree model was 69 (which is much worse).", "3392700": "Also, apart from the two mentioned paths, I believe the Training dynamics also play a significant role here. How much do you train, how do you avoid oscillations, when do you early stop, other optimizer and scheduling configurations, etc.", "3393076": "In this task, we can’t rely solely on CV or the public leaderboard, because the distributions of the train, public, and private datasets are different. A solution is more reliable if it improves both CV and the public leaderboard, and if it is a reasonable, trustworthy approach. In addition, a reliable solution should have low variance across folds.", "3393133": "now max: 71" }, "source": "meta" }