![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2405813%2Fc26f76f529598ec4418f596ff9ff09e1%2F2025-09-03%20013146.png?generation=1756859526896278&alt=media\")
Triple Grandmaster (competitions / notebooks / discussions).
\nI'm super excited - (Joined 6 years ago) and finally the MOST desirable medal!
\nI had been avoiding NN competitions for a long time, as I knew that was not good with NNs at all, but this time was different.
\nI knew exactly what I wanted to achieve and what I wanted to do with the data - I just didn't know how. Recently, I've joined JetBrains and will work on improving Junie - the agent that wrote 98% of the code for this competition - it seems that code agents will be a big part for many kagglers and upcoming competitions.
\nDidn't expect such results and super super happy!
\nNN
\nseveral heads - solo model 0.861 public
\n5 folds - sometimes different seeds and features
Full write-up make tomorrow! Promise!
\nCongrats to ALL WINNERS. Can't find words to say how much I owe to Kaggle and this community.
\nP.S. sorry for many edits - hands still shaking
", "messages": [ { "id": "3280617", "postDate": "09/03/2025 00:22:55", "content": "Triple Grandmaster (competitions / notebooks / discussions).
\nI'm super excited - (Joined 6 years ago) and finally the MOST desirable medal!
\nI had been avoiding NN competitions for a long time, as I knew that was not good with NNs at all, but this time was different.
\nI knew exactly what I wanted to achieve and what I wanted to do with the data - I just didn't know how. Recently, I've joined JetBrains and will work on improving Junie - the agent that wrote 98% of the code for this competition - it seems that code agents will be a big part for many kagglers and upcoming competitions.
\nDidn't expect such results and super super happy!
\nNN
\nseveral heads - solo model 0.861 public
\n5 folds - sometimes different seeds and features
Full write-up make tomorrow! Promise!
\nCongrats to ALL WINNERS. Can't find words to say how much I owe to Kaggle and this community.
\nP.S. sorry for many edits - hands still shaking
", "rawMarkdown": "Triple Grandmaster (competitions / notebooks / discussions).\n\nI'm super excited - (Joined 6 years ago) and finally the MOST desirable medal!\n\nI had been avoiding NN competitions for a long time, as I knew that was not good with NNs at all, but this time was different.\n\nI knew exactly what I wanted to achieve and what I wanted to do with the data - I just didn't know how. Recently, I've joined JetBrains and will work on improving Junie - the agent that wrote 98% of the code for this competition - it seems that code agents will be a big part for many kagglers and upcoming competitions.\n\nDidn't expect such results and super super happy!\n\nNN\n\n* acc (1d cnn + gru)\n* rot (1d cnn + gru)\n* velcity (1d cnn + gru)\n* custom features (1d cnn + gru)\n* thm (1d cnn + gru)\n* tof (2d cnn + gru)\n* Global aggs (MLP)\n* Demographics data (MLP)\nall in separate \"branches\"\n\nseveral heads - solo model 0.861 public\n5 folds - sometimes different seeds and features\n\n* mixup + some noise + scaling to 386 - 512 - 1024 windows\n140 base window\nseveral heads - different targets\n* ~16 models (5 folds each - many weak ones) weighted ensemble based on CV\n* Postprocess + 0.01 (std 0.002)\n\nFull write-up make tomorrow! Promise!\n\nCongrats to ALL WINNERS. Can't find words to say how much I owe to Kaggle and this community.\n\nP.S. sorry for many edits - hands still shaking", "votes": null }, { "id": "3280619", "postDate": "09/03/2025 00:26:26", "content": "Congratulations to the new GM!
", "rawMarkdown": "Congratulations to the new GM!", "votes": null }, { "id": "3280620", "postDate": "09/03/2025 00:27:32", "content": "Coding agents definitely help a lot. There are a lot of weirder or more involved ideas that I would not have bothered exploring if I didn't have something that could sketch out the code so fast and then just make minor tweaks on.
", "rawMarkdown": "Coding agents definitely help a lot. There are a lot of weirder or more involved ideas that I would not have bothered exploring if I didn't have something that could sketch out the code so fast and then just make minor tweaks on.", "votes": null }, { "id": "3280621", "postDate": "09/03/2025 00:29:03", "content": "Part of the difficulty is keeping complexity down though. If it spits out 1k lines of code very quickly and it runs, have to consider if you actually keep it or try to simplify it. By the end of a competition I accumulated a lot of crap because I didnt go through and vet everything in great detail. If I didn't have it then I would have explored much less but my codebase would be a lot cleaner. I still don't know how to keep it totally under control
", "rawMarkdown": "Part of the difficulty is keeping complexity down though. If it spits out 1k lines of code very quickly and it runs, have to consider if you actually keep it or try to simplify it. By the end of a competition I accumulated a lot of crap because I didnt go through and vet everything in great detail. If I didn't have it then I would have explored much less but my codebase would be a lot cleaner. I still don't know how to keep it totally under control", "votes": null }, { "id": "3280623", "postDate": "09/03/2025 00:31:56", "content": "This one makes a lot of sense! Congrats on your well earned medal!
", "rawMarkdown": "This one makes a lot of sense! Congrats on your well earned medal!", "votes": null }, { "id": "3280624", "postDate": "09/03/2025 00:33:04", "content": "Good prepared guidelines helps a lot. Isolated code, the style you like.
\nHe even planned experiments and executed them!
\nAnd postprocess wrote the Agent from top to bottom - unchanged till the end.
\n
A lot of automated tests.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2405813%2F0e982525d1c98a1856ad69d29432ebf9%2F2025-09-03%20013455.png?generation=1756859706825149&alt=media\")
I havent spent too much time tinkering with the configs and rules. I just seem to consistently get the 200 line variant of things more often than I get the 40 line version. Spending way too many lines on documentation, protecting against imaginary scenarios where a file or variable is missing. Often not constructing things well so a repeated process might reload a csv on every call and stuff like that.
\nIn my typical process I have a single model.py file that has all my iterations of my models and then I just import them into my train.py. This seems to completely choke the coding agent because it has many similar copies of things. Gotta keep some of these things in mind when architecting things I guess
", "rawMarkdown": "I havent spent too much time tinkering with the configs and rules. I just seem to consistently get the 200 line variant of things more often than I get the 40 line version. Spending way too many lines on documentation, protecting against imaginary scenarios where a file or variable is missing. Often not constructing things well so a repeated process might reload a csv on every call and stuff like that. \n\nIn my typical process I have a single model.py file that has all my iterations of my models and then I just import them into my train.py. This seems to completely choke the coding agent because it has many similar copies of things. Gotta keep some of these things in mind when architecting things I guess", "votes": null }, { "id": "3280628", "postDate": "09/03/2025 00:45:37", "content": "Congrats on becoming GM!! 👏🙌
", "rawMarkdown": "Congrats on becoming GM!! 👏🙌", "votes": null }, { "id": "3280632", "postDate": "09/03/2025 00:50:45", "content": "this is awesome will be great to see the agent code ; the setup looks neat . congrats on being a GM :D
", "rawMarkdown": "this is awesome will be great to see the agent code ; the setup looks neat . congrats on being a GM :D", "votes": null }, { "id": "3280641", "postDate": "09/03/2025 01:09:56", "content": "Congratulations to the new Triple Grandmaster.
", "rawMarkdown": "Congratulations to the new Triple Grandmaster.", "votes": null }, { "id": "3280643", "postDate": "09/03/2025 01:15:09", "content": "I don't have much experience about agent, but I know it's a money-cost play, would you mind reveal how much money you cost in this entire experiments?
", "rawMarkdown": "I don't have much experience about agent, but I know it's a money-cost play, would you mind reveal how much money you cost in this entire experiments?", "votes": null }, { "id": "3280644", "postDate": "09/03/2025 01:17:31", "content": "Used ultimate ai package - never got quota limit. I’m using Junie since December last year and paying 25eur month I think so.
", "rawMarkdown": "Used ultimate ai package - never got quota limit. I’m using Junie since December last year and paying 25eur month I think so.", "votes": null }, { "id": "3280648", "postDate": "09/03/2025 01:21:21", "content": "There is a “hint” - don’t use much follow-ups changes - 2-3 is ok - 10+ very costly as it’s used a lot of context (will improve this part for sure) - better rollback and start new task as there is probably a task definition issue that code agent got you thoughts in a different way (wrong way).
", "rawMarkdown": "There is a “hint” - don’t use much follow-ups changes - 2-3 is ok - 10+ very costly as it’s used a lot of context (will improve this part for sure) - better rollback and start new task as there is probably a task definition issue that code agent got you thoughts in a different way (wrong way).", "votes": null }, { "id": "3280649", "postDate": "09/03/2025 01:33:22", "content": "ahh junie that good dint know compare to copilot and like claude or does that use claude llm only under the hood
", "rawMarkdown": "ahh junie that good dint know compare to copilot and like claude or does that use claude llm only under the hood", "votes": null }, { "id": "3280651", "postDate": "09/03/2025 01:38:26", "content": "Claude / gpt5 - you choose. Other llms will appear when will pass “in-house swe benchmark” - at JetBrains it’s very strict.
", "rawMarkdown": "Claude / gpt5 - you choose. Other llms will appear when will pass “in-house swe benchmark” - at JetBrains it’s very strict.", "votes": null }, { "id": "3280667", "postDate": "09/03/2025 02:27:02", "content": "Congrats! Triple GM!!!!!
", "rawMarkdown": "Congrats! Triple GM!!!!!", "votes": null }, { "id": "3280676", "postDate": "09/03/2025 02:50:56", "content": "Congrats on the second Gold from another Child Mind Institute competition - solo at this time.
", "rawMarkdown": "Congrats on the second Gold from another Child Mind Institute competition - solo at this time.", "votes": null }, { "id": "3280686", "postDate": "09/03/2025 03:15:09", "content": "Congratulations @kyakovlev I was super excited to see you jump into gold zone public LB the past few days. And i'm happy that you finished with solo Gold private LB. Congratulations becoming triple Kaggle Grandmaster! Well deserved. You are an excellent data scientist.
", "rawMarkdown": "Congratulations @kyakovlev I was super excited to see you jump into gold zone public LB the past few days. And i'm happy that you finished with solo Gold private LB. Congratulations becoming triple Kaggle Grandmaster! Well deserved. You are an excellent data scientist.", "votes": null }, { "id": "3280701", "postDate": "09/03/2025 03:33:48", "content": "Thank you!
\nIt was an insane sprint))) was doing some models in July and didn’t even tried to submit (few test submissions to check how to deal with competition prediction process) as I thought that my models are not good enough - had the competition metric at least a point less than public notebooks on imu data. Than switched the job (JetBrains), deep dived into the new role. Suddenly, during last weekend (Friday I believe) decided to spend few free hours on things I like the most - Kaggle and found a BUG in my competition metric))))))) was checking binary f1 and just 8 classes f1 and all 9 - fixing this got two points improvement - wow I have competitive model. And now impossible task - no tof model, no post processing (was obvious from the start that is possible to use target distribution), no final imu model, just single gpu, very little free time - was madness)))
\nSo happy now)))
", "rawMarkdown": "Thank you! \n\nIt was an insane sprint))) was doing some models in July and didn’t even tried to submit (few test submissions to check how to deal with competition prediction process) as I thought that my models are not good enough - had the competition metric at least a point less than public notebooks on imu data. Than switched the job (JetBrains), deep dived into the new role. Suddenly, during last weekend (Friday I believe) decided to spend few free hours on things I like the most - Kaggle and found a BUG in my competition metric))))))) was checking binary f1 and just 8 classes f1 and all 9 - fixing this got two points improvement - wow I have competitive model. And now impossible task - no tof model, no post processing (was obvious from the start that is possible to use target distribution), no final imu model, just single gpu, very little free time - was madness))) \n\nSo happy now)))", "votes": null }, { "id": "3280722", "postDate": "09/03/2025 04:28:41", "content": "Congrats! I was super excited to see you jump into gold zone public LB the past few days.
", "rawMarkdown": "Congrats! I was super excited to see you jump into gold zone public LB the past few days.", "votes": null }, { "id": "3280725", "postDate": "09/03/2025 04:34:41", "content": "I thought the top-ranked models on the leaderboard were all grouped by subject. My score of 0.853 was based on single sequence input, so I was surprised and delighted to see 0.868—that's truly impressive.
", "rawMarkdown": "I thought the top-ranked models on the leaderboard were all grouped by subject. My score of 0.853 was based on single sequence input, so I was surprised and delighted to see 0.868—that's truly impressive.", "votes": null }, { "id": "3280727", "postDate": "09/03/2025 04:40:36", "content": "I had an idea of making sequence level model too but was too late))) 3 days before the deadline.
\nNN still hard for me - current result is luck - I don’t feel it (model), how flexible it’s, how it reacts to features, what use and where and why (yes I understand concepts but do not feel it “by skin”)
", "rawMarkdown": "I had an idea of making sequence level model too but was too late))) 3 days before the deadline.\n\nNN still hard for me - current result is luck - I don’t feel it (model), how flexible it’s, how it reacts to features, what use and where and why (yes I understand concepts but do not feel it “by skin”)", "votes": null }, { "id": "3280746", "postDate": "09/03/2025 05:19:54", "content": "Congrats! Nice Work!
\nAre you planning to start a YouTube stream called “Code with Junie”?
\nI think many Kagglers like me would be interested to join and learn about using agents in competitions!
Not me but)) but I’ll try to convince to do some educational content.
\nAs I see it - working with code agent is a partnership - you can’t expect good results if you don’t do your “homework”, like prepare instructions/define clear the task/be able to evaluate the code proposed (at least basic), correct agent if something is wrong (not just rollback and walk away) as it doesn’t understand his mistakes (everything should be converted to instructions for future use)
\nExample - code agents writes lots of annotations and comments - if you’ll just delete bad blocks the agent will continue to add extensive comments. But if you’ll add instructions to guidelines like “don’t write extensive comments - “I like this style **and small code block” he will do exactly as you want. Few iterations on errors and you have perfect partner.
", "rawMarkdown": "Not me but)) but I’ll try to convince to do some educational content. \n\nAs I see it - working with code agent is a partnership - you can’t expect good results if you don’t do your “homework”, like prepare instructions/define clear the task/be able to evaluate the code proposed (at least basic), correct agent if something is wrong (not just rollback and walk away) as it doesn’t understand his mistakes (everything should be converted to instructions for future use)\n\nExample - code agents writes lots of annotations and comments - if you’ll just delete bad blocks the agent will continue to add extensive comments. But if you’ll add instructions to guidelines like “don’t write extensive comments - “I like this style **and small code block” he will do exactly as you want. Few iterations on errors and you have perfect partner.", "votes": null }, { "id": "3280770", "postDate": "09/03/2025 06:06:53", "content": "I agree with you. I have been trying a few of the kaggle competition agents from github that score well at the mle benchmark. I think end to end solution is out. But if we can interact with the agent and intercept part of the solution, it is of great help. Human + agent seems to be feasible and strong workflow for now.
", "rawMarkdown": "I agree with you. I have been trying a few of the kaggle competition agents from github that score well at the mle benchmark. I think end to end solution is out. But if we can interact with the agent and intercept part of the solution, it is of great help. Human + agent seems to be feasible and strong workflow for now.", "votes": null }, { "id": "3280772", "postDate": "09/03/2025 06:14:21", "content": "good job! I'm also going to build full agent solutions for the next competitions.
", "rawMarkdown": "good job! I'm also going to build full agent solutions for the next competitions.", "votes": null }, { "id": "3280837", "postDate": "09/03/2025 08:18:09", "content": "Congrats…
\nAmazing Work!
Agent made PP in 30 mins - didn't even check the code due to limited time - tested on LB and worked (made few upgrades with binary predictions from one of my heads - didn't help much but I kept it)
\nExplained the idea I had and got this (more messy than normal as agent was implementing my chaotic explanations without proper instuctions)
\nhttps://www.kaggle.com/code/kyakovlev/cmi-8th-place-postprocessing
", "rawMarkdown": "Agent made PP in 30 mins - didn't even check the code due to limited time - tested on LB and worked (made few upgrades with binary predictions from one of my heads - didn't help much but I kept it)\n\nExplained the idea I had and got this (more messy than normal as agent was implementing my chaotic explanations without proper instuctions)\n\nhttps://www.kaggle.com/code/kyakovlev/cmi-8th-place-postprocessing", "votes": null }, { "id": "3280880", "postDate": "09/03/2025 09:38:29", "content": "Congrats @kyakovlev . Especially for Comp GM title. I share your happiness :)
", "rawMarkdown": "Congrats @kyakovlev . Especially for Comp GM title. I share your happiness :)", "votes": null }, { "id": "3280882", "postDate": "09/03/2025 09:43:33", "content": "Congratulations @kyakovlev, well deserved GM title!
", "rawMarkdown": "Congratulations @kyakovlev, well deserved GM title!", "votes": null }, { "id": "3280883", "postDate": "09/03/2025 09:46:11", "content": "Than you!! I’m glad that you survived shakeup also and got gold - congrats!
", "rawMarkdown": "Than you!! I’m glad that you survived shakeup also and got gold - congrats!", "votes": null }, { "id": "3286440", "postDate": "09/09/2025 20:05:36", "content": "@kyakovlev why don’t jetbeans hold a kaggle competition/hackathon for junie agent? For example if they use agent to get gold in past competitions, they will win cash or junie credits . Jetbeans end up with many kaggle templates that people can use for remix later. The community will figure out how to control this agent. Let kagglers teach kagglers !!!
", "rawMarkdown": "kyakovlev why don’t jetbeans hold a kaggle competition/hackathon for junie agent? For example if they use agent to get gold in past competitions, they will win cash or junie credits . Jetbeans end up with many kaggle templates that people can use for remix later. The community will figure out how to control this agent. Let kagglers teach kagglers !!!", "votes": null }, { "id": "3286723", "postDate": "09/10/2025 10:18:30", "content": "Like this? https://github.com/openai/mle-bench
\nI promise to think about your proposal - I feel there is something in it.
", "rawMarkdown": "Like this? https://github.com/openai/mle-bench\n\nI promise to think about your proposal - I feel there is something in it.", "votes": null }, { "id": "3286728", "postDate": "09/10/2025 10:31:02", "content": "Yes and No. the objective is not improve accuracy (current mle-bench is for modeling). I make suggesting more for marketing and application, like hackathon. The objectives:
\n1) let Junie have more template
\n2) More peple to use Junie (and hopefully locked into it and beome opensource developer for continuing improvement and subscription?)
\n3) You can have real data on how people are using Junie which is good for your product improvement
\n4) Benefit Kaggle becuase kaggle is also a agent completing platform
my idea is like this:
\nbesically you can take current vibe coding competition as reference.
\nRecent Kaggle + Nano banana AI studio is an example, it collected i think 8000 template applications. There are some bounty vibe code competition too (aim to solve vibe coded app bug)
Congratulations to the new GM!
", "votes": null, "replies": [] }, { "id": 3280620, "author_name": "ryches", "author_url": "", "post_date": "09/03/2025 00:27:32", "content": "Coding agents definitely help a lot. There are a lot of weirder or more involved ideas that I would not have bothered exploring if I didn't have something that could sketch out the code so fast and then just make minor tweaks on.
", "votes": null, "replies": [ { "id": 3280621, "author_name": "ryches", "author_url": "", "post_date": "09/03/2025 00:29:03", "content": "Part of the difficulty is keeping complexity down though. If it spits out 1k lines of code very quickly and it runs, have to consider if you actually keep it or try to simplify it. By the end of a competition I accumulated a lot of crap because I didnt go through and vet everything in great detail. If I didn't have it then I would have explored much less but my codebase would be a lot cleaner. I still don't know how to keep it totally under control
", "votes": null, "replies": [ { "id": 3280624, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 00:33:04", "content": "Good prepared guidelines helps a lot. Isolated code, the style you like.
\nHe even planned experiments and executed them!
\nAnd postprocess wrote the Agent from top to bottom - unchanged till the end.
\n
A lot of automated tests.
\n
I don't have much experience about agent, but I know it's a money-cost play, would you mind reveal how much money you cost in this entire experiments?
", "votes": null, "replies": [ { "id": 3280644, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 01:17:31", "content": "Used ultimate ai package - never got quota limit. I’m using Junie since December last year and paying 25eur month I think so.
", "votes": null, "replies": [ { "id": 3280648, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 01:21:21", "content": "There is a “hint” - don’t use much follow-ups changes - 2-3 is ok - 10+ very costly as it’s used a lot of context (will improve this part for sure) - better rollback and start new task as there is probably a task definition issue that code agent got you thoughts in a different way (wrong way).
", "votes": null, "replies": [] }, { "id": 3280649, "author_name": "gauravbrills", "author_url": "", "post_date": "09/03/2025 01:33:22", "content": "ahh junie that good dint know compare to copilot and like claude or does that use claude llm only under the hood
", "votes": null, "replies": [ { "id": 3280651, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 01:38:26", "content": "Claude / gpt5 - you choose. Other llms will appear when will pass “in-house swe benchmark” - at JetBrains it’s very strict.
", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3280627, "author_name": "ryches", "author_url": "", "post_date": "09/03/2025 00:38:41", "content": "I havent spent too much time tinkering with the configs and rules. I just seem to consistently get the 200 line variant of things more often than I get the 40 line version. Spending way too many lines on documentation, protecting against imaginary scenarios where a file or variable is missing. Often not constructing things well so a repeated process might reload a csv on every call and stuff like that.
\nIn my typical process I have a single model.py file that has all my iterations of my models and then I just import them into my train.py. This seems to completely choke the coding agent because it has many similar copies of things. Gotta keep some of these things in mind when architecting things I guess
", "votes": null, "replies": [] }, { "id": 3280632, "author_name": "gauravbrills", "author_url": "", "post_date": "09/03/2025 00:50:45", "content": "this is awesome will be great to see the agent code ; the setup looks neat . congrats on being a GM :D
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3280623, "author_name": "cody11null", "author_url": "", "post_date": "09/03/2025 00:31:56", "content": "This one makes a lot of sense! Congrats on your well earned medal!
", "votes": null, "replies": [] }, { "id": 3280628, "author_name": "urvishp80", "author_url": "", "post_date": "09/03/2025 00:45:37", "content": "Congrats on becoming GM!! 👏🙌
", "votes": null, "replies": [] }, { "id": 3280641, "author_name": "chenxin1991", "author_url": "", "post_date": "09/03/2025 01:09:56", "content": "Congratulations to the new Triple Grandmaster.
", "votes": null, "replies": [] }, { "id": 3280667, "author_name": "shanzhong8", "author_url": "", "post_date": "09/03/2025 02:27:02", "content": "Congrats! Triple GM!!!!!
", "votes": null, "replies": [] }, { "id": 3280676, "author_name": "yekenot", "author_url": "", "post_date": "09/03/2025 02:50:56", "content": "Congrats on the second Gold from another Child Mind Institute competition - solo at this time.
", "votes": null, "replies": [] }, { "id": 3280686, "author_name": "cdeotte", "author_url": "", "post_date": "09/03/2025 03:15:09", "content": "Congratulations @kyakovlev I was super excited to see you jump into gold zone public LB the past few days. And i'm happy that you finished with solo Gold private LB. Congratulations becoming triple Kaggle Grandmaster! Well deserved. You are an excellent data scientist.
", "votes": null, "replies": [ { "id": 3280701, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 03:33:48", "content": "Thank you!
\nIt was an insane sprint))) was doing some models in July and didn’t even tried to submit (few test submissions to check how to deal with competition prediction process) as I thought that my models are not good enough - had the competition metric at least a point less than public notebooks on imu data. Than switched the job (JetBrains), deep dived into the new role. Suddenly, during last weekend (Friday I believe) decided to spend few free hours on things I like the most - Kaggle and found a BUG in my competition metric))))))) was checking binary f1 and just 8 classes f1 and all 9 - fixing this got two points improvement - wow I have competitive model. And now impossible task - no tof model, no post processing (was obvious from the start that is possible to use target distribution), no final imu model, just single gpu, very little free time - was madness)))
\nSo happy now)))
", "votes": null, "replies": [] } ] }, { "id": 3280722, "author_name": "khushikyad001", "author_url": "", "post_date": "09/03/2025 04:28:41", "content": "Congrats! I was super excited to see you jump into gold zone public LB the past few days.
", "votes": null, "replies": [] }, { "id": 3280725, "author_name": "hydantess", "author_url": "", "post_date": "09/03/2025 04:34:41", "content": "I thought the top-ranked models on the leaderboard were all grouped by subject. My score of 0.853 was based on single sequence input, so I was surprised and delighted to see 0.868—that's truly impressive.
", "votes": null, "replies": [ { "id": 3280727, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 04:40:36", "content": "I had an idea of making sequence level model too but was too late))) 3 days before the deadline.
\nNN still hard for me - current result is luck - I don’t feel it (model), how flexible it’s, how it reacts to features, what use and where and why (yes I understand concepts but do not feel it “by skin”)
", "votes": null, "replies": [] } ] }, { "id": 3280746, "author_name": "hengck23", "author_url": "", "post_date": "09/03/2025 05:19:54", "content": "Congrats! Nice Work!
\nAre you planning to start a YouTube stream called “Code with Junie”?
\nI think many Kagglers like me would be interested to join and learn about using agents in competitions!
Not me but)) but I’ll try to convince to do some educational content.
\nAs I see it - working with code agent is a partnership - you can’t expect good results if you don’t do your “homework”, like prepare instructions/define clear the task/be able to evaluate the code proposed (at least basic), correct agent if something is wrong (not just rollback and walk away) as it doesn’t understand his mistakes (everything should be converted to instructions for future use)
\nExample - code agents writes lots of annotations and comments - if you’ll just delete bad blocks the agent will continue to add extensive comments. But if you’ll add instructions to guidelines like “don’t write extensive comments - “I like this style **and small code block” he will do exactly as you want. Few iterations on errors and you have perfect partner.
", "votes": null, "replies": [ { "id": 3280770, "author_name": "hengck23", "author_url": "", "post_date": "09/03/2025 06:06:53", "content": "I agree with you. I have been trying a few of the kaggle competition agents from github that score well at the mle benchmark. I think end to end solution is out. But if we can interact with the agent and intercept part of the solution, it is of great help. Human + agent seems to be feasible and strong workflow for now.
", "votes": null, "replies": [] } ] } ] }, { "id": 3280772, "author_name": "yuanzhezhou", "author_url": "", "post_date": "09/03/2025 06:14:21", "content": "good job! I'm also going to build full agent solutions for the next competitions.
", "votes": null, "replies": [] }, { "id": 3280837, "author_name": "saraharshadbcs", "author_url": "", "post_date": "09/03/2025 08:18:09", "content": "Congrats…
\nAmazing Work!
Agent made PP in 30 mins - didn't even check the code due to limited time - tested on LB and worked (made few upgrades with binary predictions from one of my heads - didn't help much but I kept it)
\nExplained the idea I had and got this (more messy than normal as agent was implementing my chaotic explanations without proper instuctions)
\nhttps://www.kaggle.com/code/kyakovlev/cmi-8th-place-postprocessing
", "votes": null, "replies": [] }, { "id": 3280880, "author_name": "davutpolat", "author_url": "", "post_date": "09/03/2025 09:38:29", "content": "Congrats @kyakovlev . Especially for Comp GM title. I share your happiness :)
", "votes": null, "replies": [] }, { "id": 3280882, "author_name": "trasibulo", "author_url": "", "post_date": "09/03/2025 09:43:33", "content": "Congratulations @kyakovlev, well deserved GM title!
", "votes": null, "replies": [ { "id": 3280883, "author_name": "kyakovlev", "author_url": "", "post_date": "09/03/2025 09:46:11", "content": "Than you!! I’m glad that you survived shakeup also and got gold - congrats!
", "votes": null, "replies": [] } ] }, { "id": 3286440, "author_name": "hengck23", "author_url": "", "post_date": "09/09/2025 20:05:36", "content": "@kyakovlev why don’t jetbeans hold a kaggle competition/hackathon for junie agent? For example if they use agent to get gold in past competitions, they will win cash or junie credits . Jetbeans end up with many kaggle templates that people can use for remix later. The community will figure out how to control this agent. Let kagglers teach kagglers !!!
", "votes": null, "replies": [ { "id": 3286723, "author_name": "kyakovlev", "author_url": "", "post_date": "09/10/2025 10:18:30", "content": "Like this? https://github.com/openai/mle-bench
\nI promise to think about your proposal - I feel there is something in it.
", "votes": null, "replies": [ { "id": 3286728, "author_name": "hengck23", "author_url": "", "post_date": "09/10/2025 10:31:02", "content": "Yes and No. the objective is not improve accuracy (current mle-bench is for modeling). I make suggesting more for marketing and application, like hackathon. The objectives:
\n1) let Junie have more template
\n2) More peple to use Junie (and hopefully locked into it and beome opensource developer for continuing improvement and subscription?)
\n3) You can have real data on how people are using Junie which is good for your product improvement
\n4) Benefit Kaggle becuase kaggle is also a agent completing platform
my idea is like this:
\nbesically you can take current vibe coding competition as reference.
\nRecent Kaggle + Nano banana AI studio is an example, it collected i think 8000 template applications. There are some bounty vibe code competition too (aim to solve vibe coded app bug)