Congrats to all the winners. Just for fun, I began working a few days before the original competition deadline. I asked ChatGPT to read the best public notebook (LB 0.552 by Tony Li here) and write updated code to improve the post process. ChatGPT's new code achieved LB 0.553 and 100th place! I was excited, but then the competition was extended another 2 weeks, haha…
\nWith an additional 2 week extension, I decided to ask ChatGPT to do more. I uploaded Tony Li's train notebook (which was Innat's training notebook here) to ChatGPT and I uploaded the host's comment here (about their uuNet baseline).
\nThen I asked ChatGPT:
\n\n\nPlease update this Jupyter notebook training pipeline to use the host's suggested model and loss. Also add deep supervision. Thank you.
\n
I begin with Tony Li's ( @tonylica ) public inference notebook here which achieved LB 0.552 (and trains TransUNet ResNext50 with 50% SparseCenterlineDiceLoss and 50% SparseDiceCELoss). The train code was provided by Innat (@ipythonx ) here. We then asked ChatGPT to add the following 5 ideas based on forum/notebook comments.
\nAsk ChatGPT to add nnUNet. ChatGPT proceeded to write a custom nnUNet from scratch in PyTorch using the host's specifications. I was amazed. See code here
\nAsk ChatGPT to add skeleton loss. ChatGPT proceeded to write an approximate of the host's MedialSurfaceRecall loss by writing a custom loss in PyTorch. I was amazed. See code here
\nAsk ChatGPT to add deep supervision. There was a comment on the public notebook about using deep supervision, so ChatGPT added deep supervision to the best public notebook too. See code here
\nThere was a comment on the public notebook about training for 1000 epochs, so we trained for 1000 epochs.
\nTo accelerate inference, I asked ChatGPT to update the code to use 2xT4 GPU during inference instead of 1xP100 GPU. See code here
\nThe original public notebook LB 0.552 had local CV Dice (on 6 hold out volumes) of 0.670. The new train code that ChatGPT wrote had CV Dice (on 6 hold out volumes) of 0.702, wow!
\nLocally, I also trained both models (a second time) using 75% data and validated with the official metric on the remaining (25%) 130 volumes. The local official metrics (using TTAx6 and PP) were 0.560 and 0.585 respectively, wow!
\nI share the code that ChatGPT wrote. Here is the train code here and here is the inference code here.
\nCongrats to all the winners. Just for fun, I began working a few days before the original competition deadline. I asked ChatGPT to read the best public notebook (LB 0.552 by Tony Li here) and write updated code to improve the post process. ChatGPT's new code achieved LB 0.553 and 100th place! I was excited, but then the competition was extended another 2 weeks, haha…
\nWith an additional 2 week extension, I decided to ask ChatGPT to do more. I uploaded Tony Li's train notebook (which was Innat's training notebook here) to ChatGPT and I uploaded the host's comment here (about their uuNet baseline).
\nThen I asked ChatGPT:
\n\n\nPlease update this Jupyter notebook training pipeline to use the host's suggested model and loss. Also add deep supervision. Thank you.
\n
I begin with Tony Li's ( @tonylica ) public inference notebook here which achieved LB 0.552 (and trains TransUNet ResNext50 with 50% SparseCenterlineDiceLoss and 50% SparseDiceCELoss). The train code was provided by Innat (@ipythonx ) here. We then asked ChatGPT to add the following 5 ideas based on forum/notebook comments.
\nAsk ChatGPT to add nnUNet. ChatGPT proceeded to write a custom nnUNet from scratch in PyTorch using the host's specifications. I was amazed. See code here
\nAsk ChatGPT to add skeleton loss. ChatGPT proceeded to write an approximate of the host's MedialSurfaceRecall loss by writing a custom loss in PyTorch. I was amazed. See code here
\nAsk ChatGPT to add deep supervision. There was a comment on the public notebook about using deep supervision, so ChatGPT added deep supervision to the best public notebook too. See code here
\nThere was a comment on the public notebook about training for 1000 epochs, so we trained for 1000 epochs.
\nTo accelerate inference, I asked ChatGPT to update the code to use 2xT4 GPU during inference instead of 1xP100 GPU. See code here
\nThe original public notebook LB 0.552 had local CV Dice (on 6 hold out volumes) of 0.670. The new train code that ChatGPT wrote had CV Dice (on 6 hold out volumes) of 0.702, wow!
\nLocally, I also trained both models (a second time) using 75% data and validated with the official metric on the remaining (25%) 130 volumes. The local official metrics (using TTAx6 and PP) were 0.560 and 0.585 respectively, wow!
\nI share the code that ChatGPT wrote. Here is the train code here and here is the inference code here.
\nFantastic results!
", "rawMarkdown": "Fantastic results!\n", "votes": 1 }, { "id": 3416802, "postDate": "2026-03-03T19:24:39.600Z", "content": "Wow, great results! 🚀\nUsing ChatGPT to redesign the training pipeline and improve CV Dice so significantly is really inspiring. Thanks for sharing the code and insights!
", "rawMarkdown": "Wow, great results! 🚀\nUsing ChatGPT to redesign the training pipeline and improve CV Dice so significantly is really inspiring. Thanks for sharing the code and insights!", "votes": 1, "replies": [ { "id": 3417460, "postDate": "2026-03-05T12:56:10.417Z", "content": "Thanks Pradeep
", "rawMarkdown": "Thanks Pradeep" } ] }, { "id": 3415348, "postDate": "2026-02-28T18:19:37.850Z", "content": "That's great
", "rawMarkdown": "That's great ", "votes": 1, "replies": [ { "id": 3415658, "postDate": "2026-03-01T04:51:15.197Z", "content": "Thanks Durga!
", "rawMarkdown": "Thanks Durga!" }, { "id": 3417710, "postDate": "2026-03-06T01:16:36.940Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] }, { "id": 3415174, "postDate": "2026-02-28T10:38:40.393Z", "content": "Haha—yeah, that’s how it is. Congratulations on your bronze medal!
\nClaude, Gemini, and ChatGPT have given me a lot of inspiration and helped me get a ton of work done.
\nI’m still a beginner. Even though I’ve had a Kaggle account for many years, I only started my competition journey this January. LLMs have helped me a lot.
\nThat said, they can’t always break out of their built-in frameworks, and they’re often very confident in their judgments. Even when I tried three different LLMs, sometimes they came up with the same solution—it looked correct, but in practice it didn’t help at all.
\nThat might be one of the main reasons we didn’t get a gold medal. LLMs can make up for some of our coding gaps, but competition experience is something I still need to build up myself.
\nAnyway, even though we got shaken out of the gold zone, I still had a great time in this competition.
\nSee you at the DPC competition!
", "rawMarkdown": "Haha—yeah, that’s how it is. Congratulations on your bronze medal!\n\nClaude, Gemini, and ChatGPT have given me a lot of inspiration and helped me get a ton of work done.\n\nI’m still a beginner. Even though I’ve had a Kaggle account for many years, I only started my competition journey this January. LLMs have helped me a lot.\n\nThat said, they can’t always break out of their built-in frameworks, and they’re often very confident in their judgments. Even when I tried three different LLMs, sometimes they came up with the same solution—it looked correct, but in practice it didn’t help at all.\n\nThat might be one of the main reasons we didn’t get a gold medal. LLMs can make up for some of our coding gaps, but competition experience is something I still need to build up myself.\n\nAnyway, even though we got shaken out of the gold zone, I still had a great time in this competition.\n\nSee you at the DPC competition!", "votes": 1, "replies": [ { "id": 3415440, "postDate": "2026-02-28T23:23:42.740Z", "content": "Congratulations @ngyzly and team. Achieving 20th place is a huge achievement in this competition! This was a difficult comp. Do not feel discouraged. Instead feel proud of your result and continue to participate in more Kaggle competitions.
\n\n\nthey can’t always break out of their built-in frameworks
\n
I agree with statement. Currently, Kagglers need to guide LLMs because LLMs cannot (yet) think out of the box to win Kaggle competitions.
", "rawMarkdown": "Congratulations @ngyzly and team. Achieving 20th place is a huge achievement in this competition! This was a difficult comp. Do not feel discouraged. Instead feel proud of your result and continue to participate in more Kaggle competitions.\n\n> they can’t always break out of their built-in frameworks\n\nI agree with statement. Currently, Kagglers need to guide LLMs because LLMs cannot (yet) think out of the box to win Kaggle competitions.\n\n" } ] }, { "id": 3415050, "postDate": "2026-02-28T05:24:29.187Z", "content": "@cdeotte Congrats.
", "rawMarkdown": "@cdeotte Congrats.", "votes": 1, "replies": [ { "id": 3415626, "postDate": "2026-03-01T04:01:42.343Z", "content": "Thanks @ipythonx Thanks for your great starter notebooks. Especially your starter train notebook! It helped me and many others get started.
", "rawMarkdown": "Thanks @ipythonx Thanks for your great starter notebooks. Especially your starter train notebook! It helped me and many others get started. ", "votes": 1 } ] }, { "id": 3415040, "postDate": "2026-02-28T05:02:19.610Z", "content": "I have a question regarding this: if everyone has access to ChatGPT, what creates a competitive edge? Given that 1,000 people might be using ChatGPT but only 100 medals are awarded, what exactly determines the final ranking among participants utilizing the same AI tool? 🤔
", "rawMarkdown": "I have a question regarding this: if everyone has access to ChatGPT, what creates a competitive edge? Given that 1,000 people might be using ChatGPT but only 100 medals are awarded, what exactly determines the final ranking among participants utilizing the same AI tool? 🤔", "votes": 1, "replies": [ { "id": 3415051, "postDate": "2026-02-28T05:26:11.923Z", "content": "if chatgpt is already kaggle grandmaster, then there is no difference in who is using it.\nbut now, chatgpt BY ITSELF, is probably a kaggler at borderline bronze or sliver level. If chatgpt is teamed up with an experienced kaggler, he can take instructions and will achieve silver and gold medals.
", "rawMarkdown": "if chatgpt is already kaggle grandmaster, then there is no difference in who is using it.\nbut now, chatgpt BY ITSELF, is probably a kaggler at borderline bronze or sliver level. If chatgpt is teamed up with an experienced kaggler, he can take instructions and will achieve silver and gold medals.", "votes": 2, "replies": [ { "id": 3415067, "postDate": "2026-02-28T05:55:55.370Z", "content": "What would happen if we created 1,000 fully autonomous bots powered by ChatGPT and let them compete against each other? What factors would determine the top 10 rankings?
", "rawMarkdown": "What would happen if we created 1,000 fully autonomous bots powered by ChatGPT and let them compete against each other? What factors would determine the top 10 rankings?", "votes": 1 }, { "id": 3415078, "postDate": "2026-02-28T06:15:54.213Z", "content": "Not good enough. So even one million agents don’t help. Agent swarm is only helpful if one agent can get solution but at moderate low probability. Now, the probability is zero.
", "rawMarkdown": "Not good enough. So even one million agents don’t help. Agent swarm is only helpful if one agent can get solution but at moderate low probability. Now, the probability is zero.", "votes": 1 } ] }, { "id": 3415076, "postDate": "2026-02-28T06:07:57.930Z", "content": "At present, coding tools need guidance from Kaggle Grandmasters to achieve Gold medal results. I have observed that they excel at writing code they are asked to write. But currently, they lack the \"think out of the box\" to achieve Gold medal results. Currently Kaggle Grandmasters are needed to guide them with ideas. Maybe someday, they will be creative enough to discover Gold by themselves.
", "rawMarkdown": "At present, coding tools need guidance from Kaggle Grandmasters to achieve Gold medal results. I have observed that they excel at writing code they are asked to write. But currently, they lack the \"think out of the box\" to achieve Gold medal results. Currently Kaggle Grandmasters are needed to guide them with ideas. Maybe someday, they will be creative enough to discover Gold by themselves.", "replies": [ { "id": 3415079, "postDate": "2026-02-28T06:16:14.700Z", "content": "Yeah, I'm referring to AI diversity. Suppose we had a single superintelligent AI — that's like local search: if you always start from the same fixed point and search the same way, you're prone to getting stuck in a local minimum. So wouldn't it be the case that even if a superintelligent AI appears in the future, we'd still need to augment it into many different versions, like humans producing many independent individuals who think differently?
", "rawMarkdown": "Yeah, I'm referring to AI diversity. Suppose we had a single superintelligent AI — that's like local search: if you always start from the same fixed point and search the same way, you're prone to getting stuck in a local minimum. So wouldn't it be the case that even if a superintelligent AI appears in the future, we'd still need to augment it into many different versions, like humans producing many independent individuals who think differently?", "votes": 1 }, { "id": 3415081, "postDate": "2026-02-28T06:24:58.310Z", "content": "That's a great point @quan0095 but I think that is \"what it means to be human\". I believe humans have \"free will\" and are therefore \"not deterministic\". Therefore multiple humans with \"free will\" do not converge to the same course of actions. The question is, \"will AI ever achieve this level of \"free will / autonomy\"?
\nCurrently humans' \"free will\" is needed to steer AI into \"out of the box\" gold medal solutions. But maybe someday AI may have \"free will\" and can find gold medal solutions themselves.
", "rawMarkdown": "That's a great point @quan0095 but I think that is \"what it means to be human\". I believe humans have \"free will\" and are therefore \"not deterministic\". Therefore multiple humans with \"free will\" do not converge to the same course of actions. The question is, \"will AI ever achieve this level of \"free will / autonomy\"?\n\nCurrently humans' \"free will\" is needed to steer AI into \"out of the box\" gold medal solutions. But maybe someday AI may have \"free will\" and can find gold medal solutions themselves.", "votes": 1 }, { "id": 3415084, "postDate": "2026-02-28T06:33:14.560Z", "content": "I also agree with that assessment \"free will\". I just think we currently only have a few specific AI tools, and in this competition — as in many others — many people will use AI tools, but not everyone who uses them will win a medal. I believe there will be a significant gap between users and non‑users of AI tools. Among those who do use AI tools, there still needs to be something that differentiates them in order to achieve a gold medal.
", "rawMarkdown": "I also agree with that assessment \"free will\". I just think we currently only have a few specific AI tools, and in this competition — as in many others — many people will use AI tools, but not everyone who uses them will win a medal. I believe there will be a significant gap between users and non‑users of AI tools. Among those who do use AI tools, there still needs to be something that differentiates them in order to achieve a gold medal.", "votes": 1 }, { "id": 3415092, "postDate": "2026-02-28T06:44:46.870Z", "content": "Why don’t you do a test. Ask the top ai coder etc to read the top 10 written solutions of this competition and then ask them to repeat results. See how well they perform. I use to train young engineers like this in my previous company
", "rawMarkdown": "Why don’t you do a test. Ask the top ai coder etc to read the top 10 written solutions of this competition and then ask them to repeat results. See how well they perform. I use to train young engineers like this in my previous company ", "votes": 3 }, { "id": 3415099, "postDate": "2026-02-28T07:01:21.393Z", "content": "One thing I like about Kaggle is that at our company we have only a small number of engineers, and no matter how talented they are the pool of ideas inevitably becomes narrow. On Kaggle, the same problem can be tackled independently by thousands of teams, producing thousands of different solutions—some of which can truly surprise us.
", "rawMarkdown": "One thing I like about Kaggle is that at our company we have only a small number of engineers, and no matter how talented they are the pool of ideas inevitably becomes narrow. On Kaggle, the same problem can be tackled independently by thousands of teams, producing thousands of different solutions—some of which can truly surprise us.", "votes": 1 }, { "id": 3415389, "postDate": "2026-02-28T20:27:05.290Z", "content": "All engineers in company should be given one month “exam or study leave” … to do kaggle and there is a kpi to get at least gold or silver medal.
", "rawMarkdown": "All engineers in company should be given one month “exam or study leave” … to do kaggle and there is a kpi to get at least gold or silver medal.", "votes": 3 } ] } ] }, { "id": 3414949, "postDate": "2026-02-28T01:46:01.537Z", "content": "Did deep supervision help? I tried it early but it didn't help me.
", "rawMarkdown": "Did deep supervision help? I tried it early but it didn't help me.", "votes": 1, "replies": [ { "id": 3414955, "postDate": "2026-02-28T01:54:23.680Z", "content": "I have limited local validation evidence. For the majority of my experiments, I only computed local dice metric (not official metric) on 6 (this is a small number) hold out volumes (so that I could submit the trained models to the LB). Then for a few experiments, I retrained on 75% volumes and computed the official metric on the remaining 25% volumes (i.e. 130 validation volumes).
\nBased on the limited validation data I have, I think using deep supervision improved local official metric by +0.002 (when training two models for the same number of fixed epochs cosine schedule)
", "rawMarkdown": "I have limited local validation evidence. For the majority of my experiments, I only computed local dice metric (not official metric) on 6 (this is a small number) hold out volumes (so that I could submit the trained models to the LB). Then for a few experiments, I retrained on 75% volumes and computed the official metric on the remaining 25% volumes (i.e. 130 validation volumes).\n\nBased on the limited validation data I have, I think using deep supervision improved local official metric by +0.002 (when training two models for the same number of fixed epochs cosine schedule)\n\n", "votes": 1 } ] }, { "id": 3414944, "postDate": "2026-02-28T01:20:55.320Z", "content": "thnaks for the writeup:\n\"The official metrics were 0.560 and 0.585 respectively, wow!\" These are local CV or public,private score?
", "rawMarkdown": "thnaks for the writeup:\n\"The official metrics were 0.560 and 0.585 respectively, wow!\" These are local CV or public,private score?", "votes": 2, "replies": [ { "id": 3414945, "postDate": "2026-02-28T01:23:46.523Z", "content": "These are computed locally by training a new model on only 75% train data and validating on the remaining 25% data. To approximate the public notebook's official metric on 25% hold out, we need to train a new local model on 75% (exactly the same as public model and stop when dice on last tf record achieves 0.670 same as public notebook's dice on the last tf record).
\nSimilarly to compute local official metric on my final submission which trained on all tf records except 1, we train a second model (with same hyperparameters) that is only trained on 75% tr records and validate on the remaining 25% tf records.
", "rawMarkdown": "These are computed locally by training a new model on only 75% train data and validating on the remaining 25% data. To approximate the public notebook's official metric on 25% hold out, we need to train a new local model on 75% (exactly the same as public model and stop when dice on last tf record achieves 0.670 same as public notebook's dice on the last tf record). \n\nSimilarly to compute local official metric on my final submission which trained on all tf records except 1, we train a second model (with same hyperparameters) that is only trained on 75% tr records and validate on the remaining 25% tf records." }, { "id": 3415028, "postDate": "2026-02-28T04:38:39.903Z", "content": "@hengck23 Is your final submission a TransUNet or uuNET? Also did you do any non-conventional (i.e. not publicly published) post process?
", "rawMarkdown": "@hengck23 Is your final submission a TransUNet or uuNET? Also did you do any non-conventional (i.e. not publicly published) post process?", "replies": [ { "id": 3415038, "postDate": "2026-02-28T05:01:09.800Z", "content": "My submission is a simple 3d residual unet with 4 layers. No of encoder blocks = 2,3,3,3. This gets public lb 0.557. Then i add refinement unet with 2,2,2,2 blocks, using concat(image, prob) as input. Two refinement rounds end up lb 0.564( which is final private 0.580, local cv 0.584 at 20% validation data). Total running time is 5+hr. All training uses deep supervision with only bce loss. No skeleton recall etc
", "rawMarkdown": "My submission is a simple 3d residual unet with 4 layers. No of encoder blocks = 2,3,3,3. This gets public lb 0.557. Then i add refinement unet with 2,2,2,2 blocks, using concat(image, prob) as input. Two refinement rounds end up lb 0.564( which is final private 0.580, local cv 0.584 at 20% validation data). Total running time is 5+hr. All training uses deep supervision with only bce loss. No skeleton recall etc", "votes": 2 }, { "id": 3415042, "postDate": "2026-02-28T05:05:54.397Z", "content": "I have another version that can recover instant surface using polynomial fitting and surface query ( detr like ). Results is good visually but bad at metric. This is because it doesn’t follow the ground truth voxel wise so it has low score for voi and surface dice. The gain in topo score is too low
", "rawMarkdown": "I have another version that can recover instant surface using polynomial fitting and surface query ( detr like ). Results is good visually but bad at metric. This is because it doesn’t follow the ground truth voxel wise so it has low score for voi and surface dice. The gain in topo score is too low", "votes": 1 }, { "id": 3415369, "postDate": "2026-02-28T19:31:02.467Z", "content": "is it possible to train your residual unet on kaggle environment?
", "rawMarkdown": "is it possible to train your residual unet on kaggle environment?" }, { "id": 3415392, "postDate": "2026-02-28T20:35:12.487Z", "content": "Hi @hengck23 great simple model (with fast train time and simple loss and architecture). I like your cascade idea of using additional models to refine predictions of previous models. I never tried doing that. It seems like a powerful technique.
", "rawMarkdown": "Hi @hengck23 great simple model (with fast train time and simple loss and architecture). I like your cascade idea of using additional models to refine predictions of previous models. I never tried doing that. It seems like a powerful technique." }, { "id": 3415410, "postDate": "2026-02-28T21:44:01.543Z", "content": "Inference code is here: https://www.kaggle.com/code/hengck23/deep-supervised-unet3d-and-2-refinement
\nfor loss function, please refer to \"model_with_loss_for_training.py\" in the code in shared dataset
", "rawMarkdown": "Inference code is here: https://www.kaggle.com/code/hengck23/deep-supervised-unet3d-and-2-refinement\n\nfor loss function, please refer to \"model_with_loss_for_training.py\" in the code in shared dataset", "votes": 2 }, { "id": 3415414, "postDate": "2026-02-28T21:51:11.067Z", "content": "So maybe you can try to add my diffeomporhic stage. It seems also can calibrate your case. Can you release the code of Polynomial fitting approach?
", "rawMarkdown": "So maybe you can try to add my [diffeomporhic stage](https://www.kaggle.com/code/tom99763/complete-sequential-pipeline-6-stages?scriptVersionId=300181765). It seems also can calibrate your case. Can you release the code of Polynomial fitting approach?", "votes": 1 } ] } ] }, { "id": 3414943, "postDate": "2026-02-28T01:20:23.410Z", "content": "Really need kaggle to add some LLM teams! That's very interesting.
", "rawMarkdown": "Really need kaggle to add some LLM teams! That's very interesting.", "votes": 2, "replies": [ { "id": 3414968, "postDate": "2026-02-28T02:32:38.253Z", "content": "complete agents as a competitors huh ıntersting
", "rawMarkdown": "complete agents as a competitors huh ıntersting ", "votes": 1 }, { "id": 3414972, "postDate": "2026-02-28T02:40:39.813Z", "content": "Yes, LLMs would do well. In the past few months, LLM coding tools have gotten much better. Now when i ask for code, it usually works on the first run. And it can write the most complicated requests easily.
", "rawMarkdown": "Yes, LLMs would do well. In the past few months, LLM coding tools have gotten much better. Now when i ask for code, it usually works on the first run. And it can write the most complicated requests easily.", "votes": 2, "replies": [ { "id": 3414978, "postDate": "2026-02-28T02:46:29.120Z", "content": "any kaggler want to finetune a code LLM on ALL historical winning notebook?
", "rawMarkdown": "any kaggler want to finetune a code LLM on ALL historical winning notebook?", "votes": 2 }, { "id": 3414984, "postDate": "2026-02-28T02:53:16.127Z", "content": "I'm not sure that finetuning is required. Current LLMs have the ability to consume large amounts of information in a conversation. So we can just give LLMs lots of relevant code (IPYNB files) and research (PDF) and ideas (URLS) in a chat discussion and it can respond with code.
", "rawMarkdown": "I'm not sure that finetuning is required. Current LLMs have the ability to consume large amounts of information in a conversation. So we can just give LLMs lots of relevant code (IPYNB files) and research (PDF) and ideas (URLS) in a chat discussion and it can respond with code." }, { "id": 3415008, "postDate": "2026-02-28T03:52:50.887Z", "content": "yeah thats one thing todo but u know what is in my mind create special agents or system prompts (or claude skills or u got the point ☺️)for specific use cases like a better one for eda that can actually constantly work on that one with capeabilities for better research like perplexity groq or gemini flash and for modelling considering more edge cases and making them work together if it isnt even yield good results at first yeah I think it is a good idea\nI am down to work with anyone about this
", "rawMarkdown": "yeah thats one thing todo but u know what is in my mind create special agents or system prompts (or claude skills or u got the point ☺️)for specific use cases like a better one for eda that can actually constantly work on that one with capeabilities for better research like perplexity groq or gemini flash and for modelling considering more edge cases and making them work together if it isnt even yield good results at first yeah I think it is a good idea\nI am down to work with anyone about this", "votes": 1 }, { "id": 3415045, "postDate": "2026-02-28T05:10:22.027Z", "content": "I am thinking if we can find a way to do this: hello chatgpt, please read ….. please output a specific md file for implementation of code so that codex can read and implement with self debug ….. we just need to find a way to write good plan md. Codex become a “super tool”
", "rawMarkdown": "I am thinking if we can find a way to do this: hello chatgpt, please read ….. please output a specific md file for implementation of code so that codex can read and implement with self debug ….. we just need to find a way to write good plan md. Codex become a “super tool”", "votes": 2 } ] }, { "id": 3415070, "postDate": "2026-02-28T05:59:09.737Z", "content": "@cdeotte do you consider use claude-code purely in tabular comp? I remember nvidia have some guide document can let it read.
", "rawMarkdown": "@cdeotte do you consider use claude-code purely in tabular comp? I remember nvidia have some guide document can let it read.", "votes": 1, "replies": [ { "id": 3415074, "postDate": "2026-02-28T06:05:34.957Z", "content": "Hi Tom, I don't understand your question. Which \"tabular comp\" are you talking about? And what \"guide document\" are you talking about?
", "rawMarkdown": "Hi Tom, I don't understand your question. Which \"tabular comp\" are you talking about? And what \"guide document\" are you talking about?", "votes": 1 }, { "id": 3415338, "postDate": "2026-02-28T17:52:16.260Z", "content": "", "rawMarkdown": "This one https://developer.nvidia.com/blog/the-kaggle-grandmasters-playbook-7-battle-tested-modeling-techniques-for-tabular-data/", "votes": 1 }, { "id": 3415363, "postDate": "2026-02-28T19:04:14.180Z", "content": "Ah ok. For the past few months, I have been trying out AI coding tools in both tabular and non-tabular competitions. Theses tools have been helping me tremendously. I believe they are game changing technology.
", "rawMarkdown": "Ah ok. For the past few months, I have been trying out AI coding tools in both tabular and non-tabular competitions. Theses tools have been helping me tremendously. I believe they are game changing technology. ", "votes": 3 } ] } ] } ], "comments": [ { "id": 3417503, "author_name": "Bhawesh Sinha", "author_url": "", "post_date": "2026-03-05T14:19:45.467000", "content": "Fantastic results!
", "votes": 1, "replies": [] }, { "id": 3416802, "author_name": "Pradeep Kumar", "author_url": "", "post_date": "2026-03-03T19:24:39.600000", "content": "Wow, great results! 🚀\nUsing ChatGPT to redesign the training pipeline and improve CV Dice so significantly is really inspiring. Thanks for sharing the code and insights!
", "votes": 1, "replies": [ { "id": 3417460, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-03-05T12:56:10.417000", "content": "Thanks Pradeep
", "votes": 0, "replies": [] } ] }, { "id": 3415348, "author_name": "Durga Kumari", "author_url": "", "post_date": "2026-02-28T18:19:37.850000", "content": "That's great
", "votes": 1, "replies": [ { "id": 3415658, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-03-01T04:51:15.197000", "content": "Thanks Durga!
", "votes": 0, "replies": [] }, { "id": 3417710, "author_name": "", "author_url": "", "post_date": "2026-03-06T01:16:36.940000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 3415174, "author_name": "Cho Royou", "author_url": "", "post_date": "2026-02-28T10:38:40.393000", "content": "Haha—yeah, that’s how it is. Congratulations on your bronze medal!
\nClaude, Gemini, and ChatGPT have given me a lot of inspiration and helped me get a ton of work done.
\nI’m still a beginner. Even though I’ve had a Kaggle account for many years, I only started my competition journey this January. LLMs have helped me a lot.
\nThat said, they can’t always break out of their built-in frameworks, and they’re often very confident in their judgments. Even when I tried three different LLMs, sometimes they came up with the same solution—it looked correct, but in practice it didn’t help at all.
\nThat might be one of the main reasons we didn’t get a gold medal. LLMs can make up for some of our coding gaps, but competition experience is something I still need to build up myself.
\nAnyway, even though we got shaken out of the gold zone, I still had a great time in this competition.
\nSee you at the DPC competition!
", "votes": 1, "replies": [ { "id": 3415440, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T23:23:42.740000", "content": "Congratulations @ngyzly and team. Achieving 20th place is a huge achievement in this competition! This was a difficult comp. Do not feel discouraged. Instead feel proud of your result and continue to participate in more Kaggle competitions.
\n\n\nthey can’t always break out of their built-in frameworks
\n
I agree with statement. Currently, Kagglers need to guide LLMs because LLMs cannot (yet) think out of the box to win Kaggle competitions.
", "votes": 0, "replies": [] } ] }, { "id": 3415050, "author_name": "Innat", "author_url": "", "post_date": "2026-02-28T05:24:29.187000", "content": "@cdeotte Congrats.
", "votes": 1, "replies": [ { "id": 3415626, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-03-01T04:01:42.343000", "content": "Thanks @ipythonx Thanks for your great starter notebooks. Especially your starter train notebook! It helped me and many others get started.
", "votes": 1, "replies": [] } ] }, { "id": 3415040, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-02-28T05:02:19.610000", "content": "I have a question regarding this: if everyone has access to ChatGPT, what creates a competitive edge? Given that 1,000 people might be using ChatGPT but only 100 medals are awarded, what exactly determines the final ranking among participants utilizing the same AI tool? 🤔
", "votes": 1, "replies": [ { "id": 3415051, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T05:26:11.923000", "content": "if chatgpt is already kaggle grandmaster, then there is no difference in who is using it.\nbut now, chatgpt BY ITSELF, is probably a kaggler at borderline bronze or sliver level. If chatgpt is teamed up with an experienced kaggler, he can take instructions and will achieve silver and gold medals.
", "votes": 2, "replies": [ { "id": 3415067, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-02-28T05:55:55.370000", "content": "What would happen if we created 1,000 fully autonomous bots powered by ChatGPT and let them compete against each other? What factors would determine the top 10 rankings?
", "votes": 1, "replies": [] }, { "id": 3415078, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T06:15:54.213000", "content": "Not good enough. So even one million agents don’t help. Agent swarm is only helpful if one agent can get solution but at moderate low probability. Now, the probability is zero.
", "votes": 1, "replies": [] } ] }, { "id": 3415076, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T06:07:57.930000", "content": "At present, coding tools need guidance from Kaggle Grandmasters to achieve Gold medal results. I have observed that they excel at writing code they are asked to write. But currently, they lack the \"think out of the box\" to achieve Gold medal results. Currently Kaggle Grandmasters are needed to guide them with ideas. Maybe someday, they will be creative enough to discover Gold by themselves.
", "votes": 0, "replies": [ { "id": 3415079, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-02-28T06:16:14.700000", "content": "Yeah, I'm referring to AI diversity. Suppose we had a single superintelligent AI — that's like local search: if you always start from the same fixed point and search the same way, you're prone to getting stuck in a local minimum. So wouldn't it be the case that even if a superintelligent AI appears in the future, we'd still need to augment it into many different versions, like humans producing many independent individuals who think differently?
", "votes": 1, "replies": [] }, { "id": 3415081, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T06:24:58.310000", "content": "That's a great point @quan0095 but I think that is \"what it means to be human\". I believe humans have \"free will\" and are therefore \"not deterministic\". Therefore multiple humans with \"free will\" do not converge to the same course of actions. The question is, \"will AI ever achieve this level of \"free will / autonomy\"?
\nCurrently humans' \"free will\" is needed to steer AI into \"out of the box\" gold medal solutions. But maybe someday AI may have \"free will\" and can find gold medal solutions themselves.
", "votes": 1, "replies": [] }, { "id": 3415084, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-02-28T06:33:14.560000", "content": "I also agree with that assessment \"free will\". I just think we currently only have a few specific AI tools, and in this competition — as in many others — many people will use AI tools, but not everyone who uses them will win a medal. I believe there will be a significant gap between users and non‑users of AI tools. Among those who do use AI tools, there still needs to be something that differentiates them in order to achieve a gold medal.
", "votes": 1, "replies": [] }, { "id": 3415092, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T06:44:46.870000", "content": "Why don’t you do a test. Ask the top ai coder etc to read the top 10 written solutions of this competition and then ask them to repeat results. See how well they perform. I use to train young engineers like this in my previous company
", "votes": 3, "replies": [] }, { "id": 3415099, "author_name": "Quan Vu", "author_url": "", "post_date": "2026-02-28T07:01:21.393000", "content": "One thing I like about Kaggle is that at our company we have only a small number of engineers, and no matter how talented they are the pool of ideas inevitably becomes narrow. On Kaggle, the same problem can be tackled independently by thousands of teams, producing thousands of different solutions—some of which can truly surprise us.
", "votes": 1, "replies": [] }, { "id": 3415389, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T20:27:05.290000", "content": "All engineers in company should be given one month “exam or study leave” … to do kaggle and there is a kpi to get at least gold or silver medal.
", "votes": 3, "replies": [] } ] } ] }, { "id": 3414949, "author_name": "CPMP", "author_url": "", "post_date": "2026-02-28T01:46:01.537000", "content": "Did deep supervision help? I tried it early but it didn't help me.
", "votes": 1, "replies": [ { "id": 3414955, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T01:54:23.680000", "content": "I have limited local validation evidence. For the majority of my experiments, I only computed local dice metric (not official metric) on 6 (this is a small number) hold out volumes (so that I could submit the trained models to the LB). Then for a few experiments, I retrained on 75% volumes and computed the official metric on the remaining 25% volumes (i.e. 130 validation volumes).
\nBased on the limited validation data I have, I think using deep supervision improved local official metric by +0.002 (when training two models for the same number of fixed epochs cosine schedule)
", "votes": 1, "replies": [] } ] }, { "id": 3414944, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T01:20:55.320000", "content": "thnaks for the writeup:\n\"The official metrics were 0.560 and 0.585 respectively, wow!\" These are local CV or public,private score?
", "votes": 2, "replies": [ { "id": 3414945, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T01:23:46.523000", "content": "These are computed locally by training a new model on only 75% train data and validating on the remaining 25% data. To approximate the public notebook's official metric on 25% hold out, we need to train a new local model on 75% (exactly the same as public model and stop when dice on last tf record achieves 0.670 same as public notebook's dice on the last tf record).
\nSimilarly to compute local official metric on my final submission which trained on all tf records except 1, we train a second model (with same hyperparameters) that is only trained on 75% tr records and validate on the remaining 25% tf records.
", "votes": 0, "replies": [] }, { "id": 3415028, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T04:38:39.903000", "content": "@hengck23 Is your final submission a TransUNet or uuNET? Also did you do any non-conventional (i.e. not publicly published) post process?
", "votes": 0, "replies": [ { "id": 3415038, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T05:01:09.800000", "content": "My submission is a simple 3d residual unet with 4 layers. No of encoder blocks = 2,3,3,3. This gets public lb 0.557. Then i add refinement unet with 2,2,2,2 blocks, using concat(image, prob) as input. Two refinement rounds end up lb 0.564( which is final private 0.580, local cv 0.584 at 20% validation data). Total running time is 5+hr. All training uses deep supervision with only bce loss. No skeleton recall etc
", "votes": 2, "replies": [] }, { "id": 3415042, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T05:05:54.397000", "content": "I have another version that can recover instant surface using polynomial fitting and surface query ( detr like ). Results is good visually but bad at metric. This is because it doesn’t follow the ground truth voxel wise so it has low score for voi and surface dice. The gain in topo score is too low
", "votes": 1, "replies": [] }, { "id": 3415369, "author_name": "MengYe", "author_url": "", "post_date": "2026-02-28T19:31:02.467000", "content": "is it possible to train your residual unet on kaggle environment?
", "votes": 0, "replies": [] }, { "id": 3415392, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T20:35:12.487000", "content": "Hi @hengck23 great simple model (with fast train time and simple loss and architecture). I like your cascade idea of using additional models to refine predictions of previous models. I never tried doing that. It seems like a powerful technique.
", "votes": 0, "replies": [] }, { "id": 3415410, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T21:44:01.543000", "content": "Inference code is here: https://www.kaggle.com/code/hengck23/deep-supervised-unet3d-and-2-refinement
\nfor loss function, please refer to \"model_with_loss_for_training.py\" in the code in shared dataset
", "votes": 2, "replies": [] }, { "id": 3415414, "author_name": "Tom", "author_url": "", "post_date": "2026-02-28T21:51:11.067000", "content": "So maybe you can try to add my diffeomporhic stage. It seems also can calibrate your case. Can you release the code of Polynomial fitting approach?
", "votes": 1, "replies": [] } ] } ] }, { "id": 3414943, "author_name": "Tom", "author_url": "", "post_date": "2026-02-28T01:20:23.410000", "content": "Really need kaggle to add some LLM teams! That's very interesting.
", "votes": 2, "replies": [ { "id": 3414968, "author_name": "Yusuf Sertkaya", "author_url": "", "post_date": "2026-02-28T02:32:38.253000", "content": "complete agents as a competitors huh ıntersting
", "votes": 1, "replies": [] }, { "id": 3414972, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T02:40:39.813000", "content": "Yes, LLMs would do well. In the past few months, LLM coding tools have gotten much better. Now when i ask for code, it usually works on the first run. And it can write the most complicated requests easily.
", "votes": 2, "replies": [ { "id": 3414978, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T02:46:29.120000", "content": "any kaggler want to finetune a code LLM on ALL historical winning notebook?
", "votes": 2, "replies": [] }, { "id": 3414984, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T02:53:16.127000", "content": "I'm not sure that finetuning is required. Current LLMs have the ability to consume large amounts of information in a conversation. So we can just give LLMs lots of relevant code (IPYNB files) and research (PDF) and ideas (URLS) in a chat discussion and it can respond with code.
", "votes": 0, "replies": [] }, { "id": 3415008, "author_name": "Yusuf Sertkaya", "author_url": "", "post_date": "2026-02-28T03:52:50.887000", "content": "yeah thats one thing todo but u know what is in my mind create special agents or system prompts (or claude skills or u got the point ☺️)for specific use cases like a better one for eda that can actually constantly work on that one with capeabilities for better research like perplexity groq or gemini flash and for modelling considering more edge cases and making them work together if it isnt even yield good results at first yeah I think it is a good idea\nI am down to work with anyone about this
", "votes": 1, "replies": [] }, { "id": 3415045, "author_name": "hengck23", "author_url": "", "post_date": "2026-02-28T05:10:22.027000", "content": "I am thinking if we can find a way to do this: hello chatgpt, please read ….. please output a specific md file for implementation of code so that codex can read and implement with self debug ….. we just need to find a way to write good plan md. Codex become a “super tool”
", "votes": 2, "replies": [] } ] }, { "id": 3415070, "author_name": "Tom", "author_url": "", "post_date": "2026-02-28T05:59:09.737000", "content": "@cdeotte do you consider use claude-code purely in tabular comp? I remember nvidia have some guide document can let it read.
", "votes": 1, "replies": [ { "id": 3415074, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T06:05:34.957000", "content": "Hi Tom, I don't understand your question. Which \"tabular comp\" are you talking about? And what \"guide document\" are you talking about?
", "votes": 1, "replies": [] }, { "id": 3415338, "author_name": "Yusuf Sertkaya", "author_url": "", "post_date": "2026-02-28T17:52:16.260000", "content": "", "votes": 1, "replies": [] }, { "id": 3415363, "author_name": "Chris Deotte", "author_url": "", "post_date": "2026-02-28T19:04:14.180000", "content": "Ah ok. For the past few months, I have been trying out AI coding tools in both tabular and non-tabular competitions. Theses tools have been helping me tremendously. I believe they are game changing technology.
", "votes": 3, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3414935": "Congrats to all the winners. Just for fun, I began working a few days before the original competition deadline. I asked ChatGPT to read the best public notebook (LB 0.552 by Tony Li [here][1]) and write updated code to improve the post process. ChatGPT's new code achieved LB 0.553 and 100th place! I was excited, but then the competition was extended another 2 weeks, haha...\n\n# ChatGPT Vibe Coding\nWith an additional 2 week extension, I decided to ask ChatGPT to do more. I uploaded Tony Li's train notebook (which was Innat's training notebook [here][3]) to ChatGPT and I uploaded the host's comment [here][2] (about their uuNet baseline). \n\nThen I asked ChatGPT:\n> Please update this Jupyter notebook training pipeline to use the host's suggested model and loss. Also add deep supervision. Thank you.\n\n# Start with Public Notebook LB 0.552\nI begin with Tony Li's ( @tonylica ) public inference notebook [here][1] which achieved LB 0.552 (and trains TransUNet ResNext50 with 50% SparseCenterlineDiceLoss and 50% SparseDiceCELoss). The train code was provided by Innat (@ipythonx ) [here][3]. We then asked ChatGPT to add the following 5 ideas based on forum/notebook comments.\n\n# => Add - Custom PyTorch nnUNet\nAsk ChatGPT to add nnUNet. ChatGPT proceeded to write a custom nnUNet from scratch in PyTorch using the host's specifications. I was amazed. See code [here][4]\n\n# => Add - Custom PyTorch Skeleton loss\nAsk ChatGPT to add skeleton loss. ChatGPT proceeded to write an approximate of the host's MedialSurfaceRecall loss by writing a custom loss in PyTorch. I was amazed. See code [here][4]\n\n# => Add - Deep Supervision\nAsk ChatGPT to add deep supervision. There was a comment on the public notebook about using deep supervision, so ChatGPT added deep supervision to the best public notebook too. See code [here][4]\n\n# => Add - 1000 epochs\nThere was a comment on the public notebook about training for 1000 epochs, so we trained for 1000 epochs.\n\n# => Add - Use 2xGPU during inference\nTo accelerate inference, I asked ChatGPT to update the code to use 2xT4 GPU during inference instead of 1xP100 GPU. See code [here][5]\n\n# Finish with Private LB 0.581 wow!\nThe original public notebook LB 0.552 had local CV Dice (on 6 hold out volumes) of 0.670. The new train code that ChatGPT wrote had CV Dice (on 6 hold out volumes) of 0.702, wow! \n\nLocally, I also trained both models (a second time) using 75% data and validated with the official metric on the remaining (25%) 130 volumes. The local official metrics (using TTAx6 and PP) were 0.560 and 0.585 respectively, wow!\n\n# Code\nI share the code that ChatGPT wrote. Here is the train code [here][4] and here is the inference code [here][5].\n\n# Enjoy!\n\n[1]: https://www.kaggle.com/code/tonylica/vesuvius-0-552?scriptVersionId=295481156\n[2]: https://www.kaggle.com/competitions/vesuvius-challenge-surface-detection/discussion/666453\n[3]: https://www.kaggle.com/code/ipythonx/train-vesuvius-surface-3d-detection-on-tpu\n[4]: https://www.kaggle.com/code/cdeotte/train-bronze-medal-uunet-by-chatgpt\n[5]: https://www.kaggle.com/code/cdeotte/infer-bronze-medal-uunet-by-chatgpt", "3417503": "Fantastic results!\n", "3416802": "Wow, great results! 🚀\nUsing ChatGPT to redesign the training pipeline and improve CV Dice so significantly is really inspiring. Thanks for sharing the code and insights!", "3415348": "That's great ", "3415174": "Haha—yeah, that’s how it is. Congratulations on your bronze medal!\n\nClaude, Gemini, and ChatGPT have given me a lot of inspiration and helped me get a ton of work done.\n\nI’m still a beginner. Even though I’ve had a Kaggle account for many years, I only started my competition journey this January. LLMs have helped me a lot.\n\nThat said, they can’t always break out of their built-in frameworks, and they’re often very confident in their judgments. Even when I tried three different LLMs, sometimes they came up with the same solution—it looked correct, but in practice it didn’t help at all.\n\nThat might be one of the main reasons we didn’t get a gold medal. LLMs can make up for some of our coding gaps, but competition experience is something I still need to build up myself.\n\nAnyway, even though we got shaken out of the gold zone, I still had a great time in this competition.\n\nSee you at the DPC competition!", "3415050": "@cdeotte Congrats.", "3415040": "I have a question regarding this: if everyone has access to ChatGPT, what creates a competitive edge? Given that 1,000 people might be using ChatGPT but only 100 medals are awarded, what exactly determines the final ranking among participants utilizing the same AI tool? 🤔", "3414949": "Did deep supervision help? I tried it early but it didn't help me.", "3414944": "thnaks for the writeup:\n\"The official metrics were 0.560 and 0.585 respectively, wow!\" These are local CV or public,private score?", "3414943": "Really need kaggle to add some LLM teams! That's very interesting." } }