i use dino v2,siglip convexnet base ,efficientnetv2 and some other i also use ndvi glcm feature heavy augmentation but i am not able to cross this score i am very frustrated again and again i got a same score i just want someone give hint idea some kind of approach that follow through the top coder because i want to learn so kindly help me if you able to help me
", "messages": [ { "id": 3378738, "postDate": "2025-12-18T16:14:54.620Z", "content": "How do you fine-tune large models like Dino V2? I'm using EfficientNet-B2 with an LB of 0.59, and I'm currently trying large models like Dino V2, but I can't fine-tune them properly. Thank you, and good luck to you.
", "rawMarkdown": "How do you fine-tune large models like Dino V2? I'm using EfficientNet-B2 with an LB of 0.59, and I'm currently trying large models like Dino V2, but I can't fine-tune them properly. Thank you, and good luck to you.", "votes": 1, "replies": [ { "id": 3378790, "postDate": "2025-12-18T16:39:24.947Z", "content": "I am only use for generating embedding and then use in xgboost and cat boost etc \nSorry I am not able to help you but when you ensemble then it cross 0.59 I know only this hopefully this is helpful for you
", "rawMarkdown": "I am only use for generating embedding and then use in xgboost and cat boost etc \nSorry I am not able to help you but when you ensemble then it cross 0.59 I know only this hopefully this is helpful for you " } ] }, { "id": 3370568, "postDate": "2025-12-10T23:49:06.183Z", "content": "any other file extracts or documentation that might indicate where the tabular features for the test set are located?
", "rawMarkdown": "any other file extracts or documentation that might indicate where the tabular features for the test set are located?", "votes": 1, "replies": [ { "id": 3370586, "postDate": "2025-12-11T00:14:14.643Z", "content": "No one has reported finding the tabular features. With data set this old Kaggle always has the risk that it's sitting in someones on-line old files :)
", "rawMarkdown": "No one has reported finding the tabular features. With data set this old Kaggle always has the risk that it's sitting in someones on-line old files :) ", "votes": 1 } ] }, { "id": 3367617, "postDate": "2025-12-08T15:31:53.013Z", "content": "In my view, the competition dataset is too small, so most models overfit easily and may achieve higher scores except for those that truly generalize well and hold top positions. This leaderboard is not final, and it’s possible that on the private leaderboard all overfitted models will drop. Your 0.61 score has a strong chance of holding a well-deserved position.
", "rawMarkdown": "In my view, the competition dataset is too small, so most models overfit easily and may achieve higher scores except for those that truly generalize well and hold top positions. This leaderboard is not final, and it’s possible that on the private leaderboard all overfitted models will drop. Your 0.61 score has a strong chance of holding a well-deserved position.", "votes": 1 }, { "id": 3367553, "postDate": "2025-12-08T14:41:38.580Z", "content": "I know this feeling @rohanrk1813. I am having the same despite my current score. Instead of providing you any approach to improve the score, I suggest you to take a step back, go through all the experiments you have done so far and analyze each. Save the training logs, Using different notebooks as that would automatically save it for you. Preferably, plot the training dynamics and compare them. Reason any performance drop and any performance gain.
\nFor example, you are using metadata. But have you observed if it is actually beneficial or not?
", "rawMarkdown": "I know this feeling @rohanrk1813. I am having the same despite my current score. Instead of providing you any approach to improve the score, I suggest you to take a step back, go through all the experiments you have done so far and analyze each. Save the training logs, Using different notebooks as that would automatically save it for you. Preferably, plot the training dynamics and compare them. Reason any performance drop and any performance gain. \n\nFor example, you are using metadata. But have you observed if it is actually beneficial or not?", "votes": 2, "replies": [ { "id": 3367616, "postDate": "2025-12-08T15:31:35.377Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3367741, "postDate": "2025-12-08T17:30:04.957Z", "content": "No i am not using metadata I extract from the image ndvi glcm color texture and it also improve my accuracy through 0.59 to 0.61
", "rawMarkdown": "No i am not using metadata I extract from the image ndvi glcm color texture and it also improve my accuracy through 0.59 to 0.61", "votes": 1 } ] }, { "id": 3367488, "postDate": "2025-12-08T13:50:04.337Z", "content": "i use dino v2,siglip convexnet base ,efficientnetv2 and some other i also use ndvi glcm feature heavy augmentation but i am not able to cross this score i am very frustrated again and again i got a same score i just want someone give hint idea some kind of approach that follow through the top coder because i want to learn so kindly help me if you able to help me
", "rawMarkdown": "i use dino v2,siglip convexnet base ,efficientnetv2 and some other i also use ndvi glcm feature heavy augmentation but i am not able to cross this score i am very frustrated again and again i got a same score i just want someone give hint idea some kind of approach that follow through the top coder because i want to learn so kindly help me if you able to help me " } ], "comments": [ { "id": 3378738, "author_name": "YYH", "author_url": "", "post_date": "2025-12-18T16:14:54.620000", "content": "How do you fine-tune large models like Dino V2? I'm using EfficientNet-B2 with an LB of 0.59, and I'm currently trying large models like Dino V2, but I can't fine-tune them properly. Thank you, and good luck to you.
", "votes": 1, "replies": [ { "id": 3378790, "author_name": "Rohanrk1813", "author_url": "", "post_date": "2025-12-18T16:39:24.947000", "content": "I am only use for generating embedding and then use in xgboost and cat boost etc \nSorry I am not able to help you but when you ensemble then it cross 0.59 I know only this hopefully this is helpful for you
", "votes": 0, "replies": [] } ] }, { "id": 3370568, "author_name": "Tarek Ziad", "author_url": "", "post_date": "2025-12-10T23:49:06.183000", "content": "any other file extracts or documentation that might indicate where the tabular features for the test set are located?
", "votes": 1, "replies": [ { "id": 3370586, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2025-12-11T00:14:14.643000", "content": "No one has reported finding the tabular features. With data set this old Kaggle always has the risk that it's sitting in someones on-line old files :)
", "votes": 1, "replies": [] } ] }, { "id": 3367617, "author_name": "Sangram Patil", "author_url": "", "post_date": "2025-12-08T15:31:53.013000", "content": "In my view, the competition dataset is too small, so most models overfit easily and may achieve higher scores except for those that truly generalize well and hold top positions. This leaderboard is not final, and it’s possible that on the private leaderboard all overfitted models will drop. Your 0.61 score has a strong chance of holding a well-deserved position.
", "votes": 1, "replies": [] }, { "id": 3367553, "author_name": "Komil Parmar", "author_url": "", "post_date": "2025-12-08T14:41:38.580000", "content": "I know this feeling @rohanrk1813. I am having the same despite my current score. Instead of providing you any approach to improve the score, I suggest you to take a step back, go through all the experiments you have done so far and analyze each. Save the training logs, Using different notebooks as that would automatically save it for you. Preferably, plot the training dynamics and compare them. Reason any performance drop and any performance gain.
\nFor example, you are using metadata. But have you observed if it is actually beneficial or not?
", "votes": 2, "replies": [ { "id": 3367616, "author_name": "", "author_url": "", "post_date": "2025-12-08T15:31:35.377000", "content": "", "votes": 0, "replies": [] }, { "id": 3367741, "author_name": "Rohanrk1813", "author_url": "", "post_date": "2025-12-08T17:30:04.957000", "content": "No i am not using metadata I extract from the image ndvi glcm color texture and it also improve my accuracy through 0.59 to 0.61
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "3378738": "How do you fine-tune large models like Dino V2? I'm using EfficientNet-B2 with an LB of 0.59, and I'm currently trying large models like Dino V2, but I can't fine-tune them properly. Thank you, and good luck to you.", "3370568": "any other file extracts or documentation that might indicate where the tabular features for the test set are located?", "3367617": "In my view, the competition dataset is too small, so most models overfit easily and may achieve higher scores except for those that truly generalize well and hold top positions. This leaderboard is not final, and it’s possible that on the private leaderboard all overfitted models will drop. Your 0.61 score has a strong chance of holding a well-deserved position.", "3367553": "I know this feeling @rohanrk1813. I am having the same despite my current score. Instead of providing you any approach to improve the score, I suggest you to take a step back, go through all the experiments you have done so far and analyze each. Save the training logs, Using different notebooks as that would automatically save it for you. Preferably, plot the training dynamics and compare them. Reason any performance drop and any performance gain. \n\nFor example, you are using metadata. But have you observed if it is actually beneficial or not?", "3367488": "i use dino v2,siglip convexnet base ,efficientnetv2 and some other i also use ndvi glcm feature heavy augmentation but i am not able to cross this score i am very frustrated again and again i got a same score i just want someone give hint idea some kind of approach that follow through the top coder because i want to learn so kindly help me if you able to help me " } }