This was a really fun competition. The dataset was a bit tiny and messy with some interesting seasonal/site distribution shifts, but this is exactly why the competition was enjoyable! I tried so many more approaches than usual because I only needed CPU compute to train my models and this was my first time really digging into what could be done with pre-trained embeddings alone. I pre-calculated all of my embeddings before training and I enjoyed trying to think about relevant augmentations that could either be reasonably pre-computed, or computed on the embeddings themselves. I spent more time than I'd like to admit fiddling with model architectures, most of which ended up being fruitless from a LB perspective, but honestly it was a great way to learn that sometimes keeping it simple and focusing on the data is really all you can or should do. I also became a lot more comfortable understanding and evaluating different model architectures/strategies (various gating strategies, mixture of experts, contrastive learning, etc.) and that alone is enough for me to walk away from this competition content with what I've learned.
\nI spent a good amount of time playing with different data splits towards the end of my effort on the competition. I'm curious how some of the approaches I tried generalize as the private leaderboard is revealed and I'll likely learn from either outcome (whether I over-fit or not). I'm looking forward to reading about some of the secret-sauce from teams that end up placing high on the leaderboard!
", "messages": [ { "id": "3396671", "postDate": "01/25/2026 14:56:56", "content": "This was a really fun competition. The dataset was a bit tiny and messy with some interesting seasonal/site distribution shifts, but this is exactly why the competition was enjoyable! I tried so many more approaches than usual because I only needed CPU compute to train my models and this was my first time really digging into what could be done with pre-trained embeddings alone. I pre-calculated all of my embeddings before training and I enjoyed trying to think about relevant augmentations that could either be reasonably pre-computed, or computed on the embeddings themselves. I spent more time than I'd like to admit fiddling with model architectures, most of which ended up being fruitless from a LB perspective, but honestly it was a great way to learn that sometimes keeping it simple and focusing on the data is really all you can or should do. I also became a lot more comfortable understanding and evaluating different model architectures/strategies (various gating strategies, mixture of experts, contrastive learning, etc.) and that alone is enough for me to walk away from this competition content with what I've learned.
\nI spent a good amount of time playing with different data splits towards the end of my effort on the competition. I'm curious how some of the approaches I tried generalize as the private leaderboard is revealed and I'll likely learn from either outcome (whether I over-fit or not). I'm looking forward to reading about some of the secret-sauce from teams that end up placing high on the leaderboard!
", "rawMarkdown": "This was a really fun competition. The dataset was a bit tiny and messy with some interesting seasonal/site distribution shifts, but this is exactly why the competition was enjoyable! I tried so many more approaches than usual because I only needed CPU compute to train my models and this was my first time really digging into what could be done with pre-trained embeddings alone. I pre-calculated all of my embeddings before training and I enjoyed trying to think about relevant augmentations that could either be reasonably pre-computed, or computed on the embeddings themselves. I spent more time than I'd like to admit fiddling with model architectures, most of which ended up being fruitless from a LB perspective, but honestly it was a great way to learn that sometimes keeping it simple and focusing on the data is really all you can or should do. I also became a lot more comfortable understanding and evaluating different model architectures/strategies (various gating strategies, mixture of experts, contrastive learning, etc.) and that alone is enough for me to walk away from this competition content with what I've learned.\n\nI spent a good amount of time playing with different data splits towards the end of my effort on the competition. I'm curious how some of the approaches I tried generalize as the private leaderboard is revealed and I'll likely learn from either outcome (whether I over-fit or not). I'm looking forward to reading about some of the secret-sauce from teams that end up placing high on the leaderboard!", "votes": null }, { "id": "3396708", "postDate": "01/25/2026 16:51:34", "content": "I don't understand the downvotes.
", "rawMarkdown": "I don't understand the downvotes.", "votes": null }, { "id": "3396736", "postDate": "01/25/2026 18:20:18", "content": "I don't either, but it didn't bother me
", "rawMarkdown": "I don't either, but it didn't bother me", "votes": null }, { "id": "3397531", "postDate": "01/27/2026 14:11:44", "content": "In my case i really dont understand that why I couldn't pass the score of single model from .71x .. somehowws I feel that there is some trick or a design difference that help the top team to surpass .73 and even .74x .. Anyways it is a good learning experience
", "rawMarkdown": "In my case i really dont understand that why I couldn't pass the score of single model from .71x .. somehowws I feel that there is some trick or a design difference that help the top team to surpass .73 and even .74x .. Anyways it is a good learning experience", "votes": null }, { "id": "3397541", "postDate": "01/27/2026 14:41:45", "content": "Yup, each architectural trick gave me only ~0.01. There is definitely a trick about the post-processing or target modeling that we haven’t found out.
", "rawMarkdown": "Yup, each architectural trick gave me only ~0.01. There is definitely a trick about the post-processing or target modeling that we haven’t found out.", "votes": null }, { "id": "3397544", "postDate": "01/27/2026 14:52:04", "content": "Can't wait to facepalm when I read the top solutions.
", "rawMarkdown": "Can't wait to facepalm when I read the top solutions.", "votes": null }, { "id": "3397566", "postDate": "01/27/2026 15:30:38", "content": "Hi, the competition says its going to end on 29th Jan 2026, does that mean its 12:00 AM in IST? Do I have final 5 submissions to make today? Or including today and tomorrow I have 10 submissions?
", "rawMarkdown": "Hi, the competition says its going to end on 29th Jan 2026, does that mean its 12:00 AM in IST? Do I have final 5 submissions to make today? Or including today and tomorrow I have 10 submissions?", "votes": null }, { "id": "3397575", "postDate": "01/27/2026 16:00:38", "content": "yes yes and it is so frustrating that ideas works on cv but not on lb 🥲🥲
", "rawMarkdown": "yes yes and it is so frustrating that ideas works on cv but not on lb 🥲🥲", "votes": null }, { "id": "3397708", "postDate": "01/27/2026 20:34:21", "content": "My best single model is .75. There is only 1 small thing separating it from other models scoring .71-.72. I do believe it is just random noise; the change doesn't make practically any difference in CV scores and I can't see any theoretical scenarios where it should have big impact. (also had to spent 10+ submissions just to pinpoint what change exactly it was causing the drop). Small training dataset + R^2 scores + different years for evaluation, I would be surprised if we don't see big shakeup tomorrow.
", "rawMarkdown": "My best single model is .75. There is only 1 small thing separating it from other models scoring .71-.72. I do believe it is just random noise; the change doesn't make practically any difference in CV scores and I can't see any theoretical scenarios where it should have big impact. (also had to spent 10+ submissions just to pinpoint what change exactly it was causing the drop). Small training dataset + R^2 scores + different years for evaluation, I would be surprised if we don't see big shakeup tomorrow.", "votes": null }, { "id": "3397871", "postDate": "01/28/2026 07:03:30", "content": "Nice rank, my best single model is 0.73, after ensemble increase to 0.74, let's see the private score tomorrow
", "rawMarkdown": "Nice rank, my best single model is 0.73, after ensemble increase to 0.74, let's see the private score tomorrow", "votes": null }, { "id": "3398205", "postDate": "01/28/2026 20:00:11", "content": "I too have a feeling I will be laughing about details I may have missed after folks start posting their submission write ups! We will see soon :)
", "rawMarkdown": "I too have a feeling I will be laughing about details I may have missed after folks start posting their submission write ups! We will see soon :)", "votes": null }, { "id": "3398206", "postDate": "01/28/2026 20:01:46", "content": "Very interesting! If it ends up generalizing, I'd love to know what that one trick was :).
", "rawMarkdown": "Very interesting! If it ends up generalizing, I'd love to know what that one trick was :).", "votes": null }, { "id": "3398209", "postDate": "01/28/2026 20:14:16", "content": "actually my best single is 0.75 and yes theres a design difference also the answer is hidden in one of the discussions only 😉 happy kaggling its been a great competiton overall fingers crossed for the private leaderboard i m damn scared🥲
", "rawMarkdown": "actually my best single is 0.75 and yes theres a design difference also the answer is hidden in one of the discussions only 😉 happy kaggling its been a great competiton overall fingers crossed for the private leaderboard i m damn scared🥲", "votes": null } ], "comments": [ { "id": 3396708, "author_name": "paritoshtripathi5", "author_url": "", "post_date": "01/25/2026 16:51:34", "content": "I don't understand the downvotes.
", "votes": null, "replies": [ { "id": 3396736, "author_name": "chemdatafarmer", "author_url": "", "post_date": "01/25/2026 18:20:18", "content": "I don't either, but it didn't bother me
", "votes": null, "replies": [] } ] }, { "id": 3397531, "author_name": "aman1391", "author_url": "", "post_date": "01/27/2026 14:11:44", "content": "In my case i really dont understand that why I couldn't pass the score of single model from .71x .. somehowws I feel that there is some trick or a design difference that help the top team to surpass .73 and even .74x .. Anyways it is a good learning experience
", "votes": null, "replies": [ { "id": 3397541, "author_name": "nlztrk", "author_url": "", "post_date": "01/27/2026 14:41:45", "content": "Yup, each architectural trick gave me only ~0.01. There is definitely a trick about the post-processing or target modeling that we haven’t found out.
", "votes": null, "replies": [ { "id": 3397544, "author_name": "tennogh", "author_url": "", "post_date": "01/27/2026 14:52:04", "content": "Can't wait to facepalm when I read the top solutions.
", "votes": null, "replies": [ { "id": 3398205, "author_name": "chemdatafarmer", "author_url": "", "post_date": "01/28/2026 20:00:11", "content": "I too have a feeling I will be laughing about details I may have missed after folks start posting their submission write ups! We will see soon :)
", "votes": null, "replies": [] } ] }, { "id": 3397575, "author_name": "aman1391", "author_url": "", "post_date": "01/27/2026 16:00:38", "content": "yes yes and it is so frustrating that ideas works on cv but not on lb 🥲🥲
", "votes": null, "replies": [] } ] }, { "id": 3397708, "author_name": "herrahuu", "author_url": "", "post_date": "01/27/2026 20:34:21", "content": "My best single model is .75. There is only 1 small thing separating it from other models scoring .71-.72. I do believe it is just random noise; the change doesn't make practically any difference in CV scores and I can't see any theoretical scenarios where it should have big impact. (also had to spent 10+ submissions just to pinpoint what change exactly it was causing the drop). Small training dataset + R^2 scores + different years for evaluation, I would be surprised if we don't see big shakeup tomorrow.
", "votes": null, "replies": [ { "id": 3397871, "author_name": "kongpasom", "author_url": "", "post_date": "01/28/2026 07:03:30", "content": "Nice rank, my best single model is 0.73, after ensemble increase to 0.74, let's see the private score tomorrow
", "votes": null, "replies": [] }, { "id": 3398206, "author_name": "chemdatafarmer", "author_url": "", "post_date": "01/28/2026 20:01:46", "content": "Very interesting! If it ends up generalizing, I'd love to know what that one trick was :).
", "votes": null, "replies": [] } ] }, { "id": 3398209, "author_name": "suman2208", "author_url": "", "post_date": "01/28/2026 20:14:16", "content": "actually my best single is 0.75 and yes theres a design difference also the answer is hidden in one of the discussions only 😉 happy kaggling its been a great competiton overall fingers crossed for the private leaderboard i m damn scared🥲
", "votes": null, "replies": [] } ] }, { "id": 3397566, "author_name": "hotsonhonet", "author_url": "", "post_date": "01/27/2026 15:30:38", "content": "Hi, the competition says its going to end on 29th Jan 2026, does that mean its 12:00 AM in IST? Do I have final 5 submissions to make today? Or including today and tomorrow I have 10 submissions?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3396671": "This was a really fun competition. The dataset was a bit tiny and messy with some interesting seasonal/site distribution shifts, but this is exactly why the competition was enjoyable! I tried so many more approaches than usual because I only needed CPU compute to train my models and this was my first time really digging into what could be done with pre-trained embeddings alone. I pre-calculated all of my embeddings before training and I enjoyed trying to think about relevant augmentations that could either be reasonably pre-computed, or computed on the embeddings themselves. I spent more time than I'd like to admit fiddling with model architectures, most of which ended up being fruitless from a LB perspective, but honestly it was a great way to learn that sometimes keeping it simple and focusing on the data is really all you can or should do. I also became a lot more comfortable understanding and evaluating different model architectures/strategies (various gating strategies, mixture of experts, contrastive learning, etc.) and that alone is enough for me to walk away from this competition content with what I've learned.\n\nI spent a good amount of time playing with different data splits towards the end of my effort on the competition. I'm curious how some of the approaches I tried generalize as the private leaderboard is revealed and I'll likely learn from either outcome (whether I over-fit or not). I'm looking forward to reading about some of the secret-sauce from teams that end up placing high on the leaderboard!", "3396708": "I don't understand the downvotes.", "3396736": "I don't either, but it didn't bother me", "3397531": "In my case i really dont understand that why I couldn't pass the score of single model from .71x .. somehowws I feel that there is some trick or a design difference that help the top team to surpass .73 and even .74x .. Anyways it is a good learning experience", "3397541": "Yup, each architectural trick gave me only ~0.01. There is definitely a trick about the post-processing or target modeling that we haven’t found out.", "3397544": "Can't wait to facepalm when I read the top solutions.", "3397566": "Hi, the competition says its going to end on 29th Jan 2026, does that mean its 12:00 AM in IST? Do I have final 5 submissions to make today? Or including today and tomorrow I have 10 submissions?", "3397575": "yes yes and it is so frustrating that ideas works on cv but not on lb 🥲🥲", "3397708": "My best single model is .75. There is only 1 small thing separating it from other models scoring .71-.72. I do believe it is just random noise; the change doesn't make practically any difference in CV scores and I can't see any theoretical scenarios where it should have big impact. (also had to spent 10+ submissions just to pinpoint what change exactly it was causing the drop). Small training dataset + R^2 scores + different years for evaluation, I would be surprised if we don't see big shakeup tomorrow.", "3397871": "Nice rank, my best single model is 0.73, after ensemble increase to 0.74, let's see the private score tomorrow", "3398205": "I too have a feeling I will be laughing about details I may have missed after folks start posting their submission write ups! We will see soon :)", "3398206": "Very interesting! If it ends up generalizing, I'd love to know what that one trick was :).", "3398209": "actually my best single is 0.75 and yes theres a design difference also the answer is hidden in one of the discussions only 😉 happy kaggling its been a great competiton overall fingers crossed for the private leaderboard i m damn scared🥲" }, "source": "meta" }