While we are at it, the most interesting bit IMHO is to estimate the winner score. It could be below 5 IMHO.
", "messages": [ { "id": 3232961, "postDate": "2025-06-26T10:58:18.707Z", "content": "While we are at it, the most interesting bit IMHO is to estimate the winner score. It could be below 5 IMHO.
", "rawMarkdown": "While we are at it, the most interesting bit IMHO is to estimate the winner score. It could be below 5 IMHO.", "votes": 8 }, { "id": 3232985, "postDate": "2025-06-26T11:30:55.427Z", "content": "It can't be estimated since @harshitsheoran has not submitted for 6 days.
\nHe may be already at below 3…
Deleted comment, hmm…
", "rawMarkdown": "Deleted comment, hmm..." } ] }, { "id": 3234234, "postDate": "2025-06-27T16:39:47.750Z", "content": "Hope for Sub 5 comes in once again, as an unexpected competitor @jeroencottaar breaks the tussle between 10-11 and enters sub 10
", "rawMarkdown": "Hope for Sub 5 comes in once again, as an unexpected competitor @jeroencottaar breaks the tussle between 10-11 and enters sub 10", "votes": 2, "replies": [ { "id": 3234439, "postDate": "2025-06-27T20:26:53.160Z", "content": "You broke it too yet said 5 is unlikely. Why would it be likely now?
\nDisclaimer, I think it is likely, which is why I posted. But i am curious about why you change your mind.
", "rawMarkdown": "You broke it too yet said 5 is unlikely. Why would it be likely now?\n\nDisclaimer, I think it is likely, which is why I posted. But i am curious about why you change your mind.", "replies": [ { "id": 3234445, "postDate": "2025-06-27T20:45:08.653Z", "content": "Well Harshit only indicated earlier that his pipeline wouldn't go sub-5. He didn't speculate about mine :)
\nThat being said, mine is also not going sub 5 without some major breakthrough.
", "rawMarkdown": "Well Harshit only indicated earlier that *his* pipeline wouldn't go sub-5. He didn't speculate about mine :)\n\nThat being said, mine is also not going sub 5 without some major breakthrough.", "votes": 3 }, { "id": 3234806, "postDate": "2025-06-28T10:53:44.613Z", "content": "Hello! @cpmpml ,
\nI didn't change my mind about my pipeline, not getting a sub 5 with that, but I think I can see sub 5 being possible, will take much more time than what's left in the comp… but I can see it happening.
\nMy hope was with maybe @jeroencottaar has a much better pipeline as he went from 14.x to 9.8 in a single day with 2 submissions
Thank you both. Let's see how far people can go in the comp.
", "rawMarkdown": "Thank you both. Let's see how far people can go in the comp." }, { "id": 3234835, "postDate": "2025-06-28T11:44:03.820Z", "content": "\n\nHello! @cpmpml ,
\n
\n I didn't change my mind about my pipeline, not getting a sub 5 with that, but I think I can see sub 5 being possible, will take much more time than what's left in the comp… but I can see it happening.
\n My hope was with maybe @jeroencottaar has a much better pipeline as he went from 14.x to 9.8 in a single day with 2 submissions
Just hadn't submitted in a while, the gain came over a much longer period.
", "rawMarkdown": "> Hello! @cpmpml ,\n> I didn't change my mind about my pipeline, not getting a sub 5 with that, but I think I can see sub 5 being possible, will take much more time than what's left in the comp... but I can see it happening. \n> My hope was with maybe @jeroencottaar has a much better pipeline as he went from 14.x to 9.8 in a single day with 2 submissions\n\nJust hadn't submitted in a while, the gain came over a much longer period.", "votes": 2 } ] }, { "id": 3234481, "postDate": "2025-06-27T22:33:46.093Z", "content": "i would say it is a matter of time and compute. just by repeating the current methods, loss can still can go down … more slowly.
\none of my first kaggle competitions is about car segmentation. i just note that one car in test is not in training set. so i just use heurstics to increase the missing segmentation (e.g the segmentation prediction must at least be continous and have car shape). this is not machine learning for that particular car, but simply heurstics … and it works! (becuase all ML methods of other team would failed for that car.
\nmost people don't look at the results and analyse source of errors. there could be some easy improvement…
", "rawMarkdown": "i would say it is a matter of time and compute. just by repeating the current methods, loss can still can go down ... more slowly.\n\none of my first kaggle competitions is about car segmentation. i just note that one car in test is not in training set. so i just use heurstics to increase the missing segmentation (e.g the segmentation prediction must at least be continous and have car shape). this is not machine learning for that particular car, but simply heurstics ... and it works! (becuase all ML methods of other team would failed for that car.\n\nmost people don't look at the results and analyse source of errors. there could be some easy improvement...", "replies": [ { "id": 3234484, "postDate": "2025-06-27T22:43:30.500Z", "content": "Your team's ranking-up speed is remarkably fast, and it seems like you still haven't encountered any obstacles.
", "rawMarkdown": "Your team's ranking-up speed is remarkably fast, and it seems like you still haven't encountered any obstacles." }, { "id": 3234791, "postDate": "2025-06-28T10:19:57.917Z", "content": "I am a late team member in this competition. Haven't contributed much.
\nHuge credits must go the our team leader and the rest of the team members for the very good results so far !!
\nThey have very great ideas and solid implementation.
Your team did an amazing job!
", "rawMarkdown": "Your team did an amazing job!" } ] } ] }, { "id": 3233028, "postDate": "2025-06-26T12:32:12.723Z", "content": "With the method/pipeline I use, a sub 5 score will need way too much compute and time, I don't think I have it even close.
", "rawMarkdown": "With the method/pipeline I use, a sub 5 score will need way too much compute and time, I don't think I have it even close.", "votes": 1, "replies": [ { "id": 3233032, "postDate": "2025-06-26T12:35:20.393Z", "content": "If you say you can't, I will guess no one else will
", "rawMarkdown": "If you say you can't, I will guess no one else will" }, { "id": 3233113, "postDate": "2025-06-26T13:42:55.757Z", "content": "Maybe greySnow is hiding his best submission, just to catch @harshitsheoran off guard :)
", "rawMarkdown": "Maybe greySnow is hiding his best submission, just to catch @harshitsheoran off guard :)", "replies": [ { "id": 3233126, "postDate": "2025-06-26T13:48:11.190Z", "content": "I have nothing to hide, my current sub is my best sub to date lol
", "rawMarkdown": "I have nothing to hide, my current sub is my best sub to date lol", "votes": 2 }, { "id": 3233145, "postDate": "2025-06-26T13:56:10.183Z", "content": "I believe that a better solution should win, even I want to see something like a sub 5 LB even if it's someone else that does it :D
", "rawMarkdown": "I believe that a better solution should win, even I want to see something like a sub 5 LB even if it's someone else that does it :D", "votes": 1 } ] } ] }, { "id": 3235428, "postDate": "2025-06-29T08:32:18.863Z", "content": "Where can we bet on it?
", "rawMarkdown": "Where can we bet on it?" }, { "id": 3233909, "postDate": "2025-06-27T10:13:23.557Z", "content": "Would you mind sharing one of your training/validation curves?
", "rawMarkdown": "Would you mind sharing one of your training/validation curves?", "replies": [ { "id": 3234156, "postDate": "2025-06-27T15:02:58.463Z", "content": "I won't share anything before competition end.
\nAnd I don't use any curve. I prefer to look at raw numbers.
", "rawMarkdown": "I won't share anything before competition end.\n\nAnd I don't use any curve. I prefer to look at raw numbers.", "votes": 3 } ] }, { "id": 3233827, "postDate": "2025-06-27T08:06:08.380Z", "content": "Is there a risk of overfitting to the public leaderboard? It seems like everyone is racing to optimize for it.
", "rawMarkdown": "Is there a risk of overfitting to the public leaderboard? It seems like everyone is racing to optimize for it.", "replies": [ { "id": 3234070, "postDate": "2025-06-27T13:34:28.357Z", "content": "The volume of the data is so large that it's not really possible to overfit the public leaderboard. The only possible concern could be a population shift between the public and private leaderboard, but this is not typically present.
", "rawMarkdown": "The volume of the data is so large that it's not really possible to overfit the public leaderboard. The only possible concern could be a population shift between the public and private leaderboard, but this is not typically present.", "votes": 1, "replies": [ { "id": 3234157, "postDate": "2025-06-27T15:03:42.923Z", "content": "We all hope there is no distribution shift between public and private test data…
", "rawMarkdown": "We all hope there is no distribution shift between public and private test data...", "votes": 3 } ] } ] }, { "id": 3233112, "postDate": "2025-06-26T13:39:17.243Z", "content": "\n\nis to estimate the winner score. It could be below 5.
\n
Sure, but do you mean the public or private LB score?
", "rawMarkdown": ">is to estimate the winner score. It could be below 5.\n\nSure, but do you mean the public or private LB score?", "replies": [ { "id": 3233160, "postDate": "2025-06-26T14:16:10.547Z", "content": "The winner score is the best score on private LB. Public LB has no value at the end of competition.
", "rawMarkdown": "The winner score is the best score on private LB. Public LB has no value at the end of competition." } ] } ], "comments": [ { "id": 3232985, "author_name": "greySnow", "author_url": "", "post_date": "2025-06-26T11:30:55.427000", "content": "It can't be estimated since @harshitsheoran has not submitted for 6 days.
\nHe may be already at below 3…
Deleted comment, hmm…
", "votes": 0, "replies": [] } ] }, { "id": 3234234, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-27T16:39:47.750000", "content": "Hope for Sub 5 comes in once again, as an unexpected competitor @jeroencottaar breaks the tussle between 10-11 and enters sub 10
", "votes": 2, "replies": [ { "id": 3234439, "author_name": "CPMP", "author_url": "", "post_date": "2025-06-27T20:26:53.160000", "content": "You broke it too yet said 5 is unlikely. Why would it be likely now?
\nDisclaimer, I think it is likely, which is why I posted. But i am curious about why you change your mind.
", "votes": 0, "replies": [ { "id": 3234445, "author_name": "Jeroen Cottaar", "author_url": "", "post_date": "2025-06-27T20:45:08.653000", "content": "Well Harshit only indicated earlier that his pipeline wouldn't go sub-5. He didn't speculate about mine :)
\nThat being said, mine is also not going sub 5 without some major breakthrough.
", "votes": 3, "replies": [] }, { "id": 3234806, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-28T10:53:44.613000", "content": "Hello! @cpmpml ,
\nI didn't change my mind about my pipeline, not getting a sub 5 with that, but I think I can see sub 5 being possible, will take much more time than what's left in the comp… but I can see it happening.
\nMy hope was with maybe @jeroencottaar has a much better pipeline as he went from 14.x to 9.8 in a single day with 2 submissions
Thank you both. Let's see how far people can go in the comp.
", "votes": 0, "replies": [] }, { "id": 3234835, "author_name": "Jeroen Cottaar", "author_url": "", "post_date": "2025-06-28T11:44:03.820000", "content": "\n\nHello! @cpmpml ,
\n
\n I didn't change my mind about my pipeline, not getting a sub 5 with that, but I think I can see sub 5 being possible, will take much more time than what's left in the comp… but I can see it happening.
\n My hope was with maybe @jeroencottaar has a much better pipeline as he went from 14.x to 9.8 in a single day with 2 submissions
Just hadn't submitted in a while, the gain came over a much longer period.
", "votes": 2, "replies": [] } ] }, { "id": 3234481, "author_name": "hengck23", "author_url": "", "post_date": "2025-06-27T22:33:46.093000", "content": "i would say it is a matter of time and compute. just by repeating the current methods, loss can still can go down … more slowly.
\none of my first kaggle competitions is about car segmentation. i just note that one car in test is not in training set. so i just use heurstics to increase the missing segmentation (e.g the segmentation prediction must at least be continous and have car shape). this is not machine learning for that particular car, but simply heurstics … and it works! (becuase all ML methods of other team would failed for that car.
\nmost people don't look at the results and analyse source of errors. there could be some easy improvement…
", "votes": 0, "replies": [ { "id": 3234484, "author_name": "guo dashuai", "author_url": "", "post_date": "2025-06-27T22:43:30.500000", "content": "Your team's ranking-up speed is remarkably fast, and it seems like you still haven't encountered any obstacles.
", "votes": 0, "replies": [] }, { "id": 3234791, "author_name": "hengck23", "author_url": "", "post_date": "2025-06-28T10:19:57.917000", "content": "I am a late team member in this competition. Haven't contributed much.
\nHuge credits must go the our team leader and the rest of the team members for the very good results so far !!
\nThey have very great ideas and solid implementation.
Your team did an amazing job!
", "votes": 0, "replies": [] } ] } ] }, { "id": 3233028, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-26T12:32:12.723000", "content": "With the method/pipeline I use, a sub 5 score will need way too much compute and time, I don't think I have it even close.
", "votes": 1, "replies": [ { "id": 3233032, "author_name": "greySnow", "author_url": "", "post_date": "2025-06-26T12:35:20.393000", "content": "If you say you can't, I will guess no one else will
", "votes": 0, "replies": [] }, { "id": 3233113, "author_name": "c-number", "author_url": "", "post_date": "2025-06-26T13:42:55.757000", "content": "Maybe greySnow is hiding his best submission, just to catch @harshitsheoran off guard :)
", "votes": 0, "replies": [ { "id": 3233126, "author_name": "greySnow", "author_url": "", "post_date": "2025-06-26T13:48:11.190000", "content": "I have nothing to hide, my current sub is my best sub to date lol
", "votes": 2, "replies": [] }, { "id": 3233145, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2025-06-26T13:56:10.183000", "content": "I believe that a better solution should win, even I want to see something like a sub 5 LB even if it's someone else that does it :D
", "votes": 1, "replies": [] } ] } ] }, { "id": 3235428, "author_name": "Man of the year", "author_url": "", "post_date": "2025-06-29T08:32:18.863000", "content": "Where can we bet on it?
", "votes": 0, "replies": [] }, { "id": 3233909, "author_name": "Cyrus", "author_url": "", "post_date": "2025-06-27T10:13:23.557000", "content": "Would you mind sharing one of your training/validation curves?
", "votes": 0, "replies": [ { "id": 3234156, "author_name": "CPMP", "author_url": "", "post_date": "2025-06-27T15:02:58.463000", "content": "I won't share anything before competition end.
\nAnd I don't use any curve. I prefer to look at raw numbers.
", "votes": 3, "replies": [] } ] }, { "id": 3233827, "author_name": "Cyrus", "author_url": "", "post_date": "2025-06-27T08:06:08.380000", "content": "Is there a risk of overfitting to the public leaderboard? It seems like everyone is racing to optimize for it.
", "votes": 0, "replies": [ { "id": 3234070, "author_name": "Jeroen Cottaar", "author_url": "", "post_date": "2025-06-27T13:34:28.357000", "content": "The volume of the data is so large that it's not really possible to overfit the public leaderboard. The only possible concern could be a population shift between the public and private leaderboard, but this is not typically present.
", "votes": 1, "replies": [ { "id": 3234157, "author_name": "CPMP", "author_url": "", "post_date": "2025-06-27T15:03:42.923000", "content": "We all hope there is no distribution shift between public and private test data…
", "votes": 3, "replies": [] } ] } ] }, { "id": 3233112, "author_name": "Truth Seeker", "author_url": "", "post_date": "2025-06-26T13:39:17.243000", "content": "\n\nis to estimate the winner score. It could be below 5.
\n
Sure, but do you mean the public or private LB score?
", "votes": 0, "replies": [ { "id": 3233160, "author_name": "CPMP", "author_url": "", "post_date": "2025-06-26T14:16:10.547000", "content": "The winner score is the best score on private LB. Public LB has no value at the end of competition.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3232961": "While we are at it, the most interesting bit IMHO is to estimate the winner score. It could be below 5 IMHO.", "3232985": "It can't be estimated since @harshitsheoran has not submitted for 6 days. \nHe may be already at below 3...", "3234234": "Hope for Sub 5 comes in once again, as an unexpected competitor @jeroencottaar breaks the tussle between 10-11 and enters sub 10", "3233028": "With the method/pipeline I use, a sub 5 score will need way too much compute and time, I don't think I have it even close.", "3235428": "Where can we bet on it?", "3233909": "Would you mind sharing one of your training/validation curves?", "3233827": "Is there a risk of overfitting to the public leaderboard? It seems like everyone is racing to optimize for it.", "3233112": ">is to estimate the winner score. It could be below 5.\n\nSure, but do you mean the public or private LB score?" } }