![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F768166%2F16e5c5c68fc1ce2af38a88d4d5c0b43e%2FScreenshot%202026-01-05%20at%204.06.20PM.png?generation=1767647222013611&alt=media\")
and that is the predicimant in the comp I'm curious if other folks have run into this. Some architectures I've played with are pretty stable and regardless of notebook environment and seeds used, you get a pretty stable model.
\nOthers seem really sensitive. I'm talking about massive differences of either fitting the train data well (leaderboard is a different story), or regression straight to the mean and getting stuck.
\nI'm seeding everything I can think of (Keras, Jax, Numpy, etc.), but even then some models seem brittle. I'm curious if other folks have run into this with their models.
\nLooking at you Genius Seed is all you need, haha.
", "messages": [ { "id": "3385206", "postDate": "01/02/2026 17:55:11", "content": "I'm curious if other folks have run into this. Some architectures I've played with are pretty stable and regardless of notebook environment and seeds used, you get a pretty stable model.
\nOthers seem really sensitive. I'm talking about massive differences of either fitting the train data well (leaderboard is a different story), or regression straight to the mean and getting stuck.
\nI'm seeding everything I can think of (Keras, Jax, Numpy, etc.), but even then some models seem brittle. I'm curious if other folks have run into this with their models.
\nLooking at you Genius Seed is all you need, haha.
", "rawMarkdown": "I'm curious if other folks have run into this. Some architectures I've played with are pretty stable and regardless of notebook environment and seeds used, you get a pretty stable model.\n\nOthers seem really sensitive. I'm talking about massive differences of either fitting the train data well (leaderboard is a different story), or regression straight to the mean and getting stuck.\n\nI'm seeding everything I can think of (Keras, Jax, Numpy, etc.), but even then some models seem brittle. I'm curious if other folks have run into this with their models.\n\nLooking at you Genius Seed is all you need, haha.", "votes": null }, { "id": "3385216", "postDate": "01/02/2026 18:19:37", "content": "IMO the only reason to spend lots of hours/days on this set of data is to learn tools that you need when you have a really bad small data set :)
\nSince it's small you can run a whole lot of ideas thru your compute. For sure seeding has played a role in the past but like you I found nothing yet. But still having fun for sure.
", "rawMarkdown": "IMO the only reason to spend lots of hours/days on this set of data is to learn tools that you need when you have a really bad small data set :)\n\nSince it's small you can run a whole lot of ideas thru your compute. For sure seeding has played a role in the past but like you I found nothing yet. But still having fun for sure.", "votes": null }, { "id": "3385225", "postDate": "01/02/2026 18:42:46", "content": "I think that's a perfectly valid perspective! I'm glad you're also having fun. I have really enjoyed this competition specifically because the data set is so small, which makes iterating super easy. Heck, I'm training many of my models on CPU compute alone!
\nI do think there is signal here though. Without taking any leaderboard behavior into account, I can get between 0.65-0.69 score fairly consistently (as have many others). Given the data set size and cleanliness, as well as the split chosen for the leaderboard, I actually think that's pretty good!
", "rawMarkdown": "I think that's a perfectly valid perspective! I'm glad you're also having fun. I have really enjoyed this competition specifically because the data set is so small, which makes iterating super easy. Heck, I'm training many of my models on CPU compute alone!\n\nI do think there is signal here though. Without taking any leaderboard behavior into account, I can get between 0.65-0.69 score fairly consistently (as have many others). Given the data set size and cleanliness, as well as the split chosen for the leaderboard, I actually think that's pretty good!", "votes": null }, { "id": "3386794", "postDate": "01/05/2026 21:07:27", "content": "yes true have found some trainings smooth and some are topsy turvy .. for some it like spikes and then flattens also and that is the predicimant in the comp
IMO the only reason to spend lots of hours/days on this set of data is to learn tools that you need when you have a really bad small data set :)
\nSince it's small you can run a whole lot of ideas thru your compute. For sure seeding has played a role in the past but like you I found nothing yet. But still having fun for sure.
", "votes": null, "replies": [ { "id": 3385225, "author_name": "chemdatafarmer", "author_url": "", "post_date": "01/02/2026 18:42:46", "content": "I think that's a perfectly valid perspective! I'm glad you're also having fun. I have really enjoyed this competition specifically because the data set is so small, which makes iterating super easy. Heck, I'm training many of my models on CPU compute alone!
\nI do think there is signal here though. Without taking any leaderboard behavior into account, I can get between 0.65-0.69 score fairly consistently (as have many others). Given the data set size and cleanliness, as well as the split chosen for the leaderboard, I actually think that's pretty good!
", "votes": null, "replies": [] } ] }, { "id": 3386794, "author_name": "gauravbrills", "author_url": "", "post_date": "01/05/2026 21:07:27", "content": "yes true have found some trainings smooth and some are topsy turvy .. for some it like spikes and then flattens also and that is the predicimant in the comp