![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2832287%2F8a809d9e8c4d657c631db295afda0c25%2FScreenshot%202026-01-19%20at%2013.11.14.png?generation=1768803103667605&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2832287%2F5e22a1b48e1d7d4b40352110ff1e24c0%2FScreenshot%202026-01-19%20at%2013.12.21.png?generation=1768803174037558&alt=media\")
In competitions with small datasets (like this one with ~350 unique training images), the Public LB is calculated on maybe 20-30% of the test data. That means the leaderboard score is based on just ~50-100 images.\nIf you tune your model to improve the LB score on those 50 images, you are overfitting the leaderboard.\nWhen the Private LB (the other 70%) is revealed at the end, models that chased the Public LB usually crash. Models with robust CV rise to the top.
\nThus, I think we need to focus on improving CV
", "messages": [ { "id": "3393490", "postDate": "01/19/2026 06:10:06", "content": "In competitions with small datasets (like this one with ~350 unique training images), the Public LB is calculated on maybe 20-30% of the test data. That means the leaderboard score is based on just ~50-100 images.\nIf you tune your model to improve the LB score on those 50 images, you are overfitting the leaderboard.\nWhen the Private LB (the other 70%) is revealed at the end, models that chased the Public LB usually crash. Models with robust CV rise to the top.
\nThus, I think we need to focus on improving CV
", "rawMarkdown": "In competitions with small datasets (like this one with ~350 unique training images), the Public LB is calculated on maybe 20-30% of the test data. That means the leaderboard score is based on just ~50-100 images.\nIf you tune your model to improve the LB score on those 50 images, you are overfitting the leaderboard.\nWhen the Private LB (the other 70%) is revealed at the end, models that chased the Public LB usually crash. Models with robust CV rise to the top.\n\nThus, I think we need to focus on improving CV", "votes": null }, { "id": "3393492", "postDate": "01/19/2026 06:13:02", "content": "\n
I think that LB is more important. CV seems to be a bit leaky.
", "rawMarkdown": "I think that LB is more important. CV seems to be a bit leaky.", "votes": null }, { "id": "3393747", "postDate": "01/19/2026 17:36:56", "content": "Yeah, you are right, I neglected the info. But I think both 357 and 800 images are few to build better model that generalizes well.
", "rawMarkdown": "Yeah, you are right, I neglected the info. But I think both 357 and 800 images are few to build better model that generalizes well.", "votes": null }, { "id": "3393748", "postDate": "01/19/2026 17:37:45", "content": "You should watch out for the leakage.
", "rawMarkdown": "You should watch out for the leakage.", "votes": null } ], "comments": [ { "id": 3393492, "author_name": "quan0095", "author_url": "", "post_date": "01/19/2026 06:13:02", "content": "\n
Yeah, you are right, I neglected the info. But I think both 357 and 800 images are few to build better model that generalizes well.
", "votes": null, "replies": [] } ] }, { "id": 3393731, "author_name": "jankowalski2000", "author_url": "", "post_date": "01/19/2026 17:02:34", "content": "I think that LB is more important. CV seems to be a bit leaky.
", "votes": null, "replies": [ { "id": 3393748, "author_name": "ibratusmonov", "author_url": "", "post_date": "01/19/2026 17:37:45", "content": "You should watch out for the leakage.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3393490": "In competitions with small datasets (like this one with ~350 unique training images), the Public LB is calculated on maybe 20-30% of the test data. That means the leaderboard score is based on just ~50-100 images.\nIf you tune your model to improve the LB score on those 50 images, you are overfitting the leaderboard.\nWhen the Private LB (the other 70%) is revealed at the end, models that chased the Public LB usually crash. Models with robust CV rise to the top.\n\nThus, I think we need to focus on improving CV", "3393492": "\n", "3393731": "I think that LB is more important. CV seems to be a bit leaky.", "3393747": "Yeah, you are right, I neglected the info. But I think both 357 and 800 images are few to build better model that generalizes well.", "3393748": "You should watch out for the leakage." }, "source": "meta" }