In the public leader board i can see around 350 people stuck at 35.1 score. I have trained my model with 77k rows high quality self made dataset, written massive codes to extract data from different pdf , book , website etc which took me 2.5 weeks still i am at 34 idk is there something i am missing or 800 people above are also collecting that much data like akkadian data is not easy to find and extract
\nif you see the board under 34 score at increase of 1 point you will mostly cross 40-50 teams but here between 34 and 35 you are crossing 700 teamssss , i dont have any words The score is accommodated between 34 and 35 how is that possible???????/
\nplease someone tell me what is happening up there
", "messages": [ { "id": 3408489, "postDate": "2026-02-20T18:18:17.640Z", "content": "All those people are just unoriginal forkers, people who are not truly interested in competing and rather copy/paste other people's best scoring public notebooks and submit them. That is why you see so many people with the same score. Don't worry, those public notebooks almost always are heavily overfitting the public leaderboard so in the end they get all shaked down.
", "rawMarkdown": "All those people are just unoriginal forkers, people who are not truly interested in competing and rather copy/paste other people's best scoring public notebooks and submit them. That is why you see so many people with the same score. Don't worry, those public notebooks almost always are heavily overfitting the public leaderboard so in the end they get all shaked down.", "votes": 7, "replies": [ { "id": 3408543, "postDate": "2026-02-20T20:03:46.960Z", "content": "exactly this
", "rawMarkdown": "exactly this" }, { "id": 3408613, "postDate": "2026-02-21T01:01:14.603Z", "content": "Feeling good after hearing that
\nBut there should be some moderation over what is submitted like the people who are genuinely doing the hard work are betten buy someone who is just copy and pasting model no real work
", "rawMarkdown": "Feeling good after hearing that\n\nBut there should be some moderation over what is submitted like the people who are genuinely doing the hard work are betten buy someone who is just copy and pasting model no real work", "votes": 1, "replies": [ { "id": 3410547, "postDate": "2026-02-23T08:52:35.053Z", "content": "This post shows how grim the situation is right now (https://www.kaggle.com/discussions/product-feedback/671127#3400122) and it feels like Kaggle doesn't really care… moth said public LB toppers \"almost always\" overfit, but honestly, that hasn't happened in most of the comps I've joined lately.
\nThere's a lot of medal corruption going on. Even some gold medal write-ups basically say \"oh I admit just merged the highest scoring public model in to my own and got lucky lol.\" From my experience, out of 5 competitions in the last year, only one actually had a real shake down for the copy-cats. If you go into a competition expecting a fair leaderboard, you’re gonna be disappointed when public notebook forkers fill up the silver/bronze tiers at the end. The reality is the leaderboard (outside of the top few genuinely innovative gold notebooks) is mostly just based on who forked the highest score kernel fastest and who got lucky with a random seed. That’s just a fact at this point.
\nAlso, many gold solutions at the top don't come from a standard 16gb consumer gpu. When the write ups come out, you will see massive ensembles that required burning hundreds of gpu hours (I am not denying building at huge scale also requires an expertise, having access to hardware capacity isn't equal to win ofcourse but if we are talking about fairness lets put it down). They are not usually something a typical home user can do without spending a fortune on cloud costs. If we’re lucky, maybe 1 or 2 real experts actually beat the game with a clever approach that fits on a normal scale, and we get to read and learn from them. Unless you are real expert in field so that you can beat the scale with innovation, IMO the only reason to even attend a competition now is for the learning opportunity and getting your hands dirty on a new subject. Expecting a fair rank for hard work is just asking for a headache. Sorry to hear that you spent alot of time doing the boring data extraction & cleaning staff, lets hope there will be a shakedown and genuinely hard work like yours gets rewarded this time. 🍀
", "rawMarkdown": "This post shows how grim the situation is right now (https://www.kaggle.com/discussions/product-feedback/671127#3400122) and it feels like Kaggle doesn't really care... moth said public LB toppers \"almost always\" overfit, but honestly, that hasn't happened in most of the comps I've joined lately.\n\nThere's a lot of medal corruption going on. Even some gold medal write-ups basically say \"oh I admit just merged the highest scoring public model in to my own and got lucky lol.\" From my experience, out of 5 competitions in the last year, only one actually had a real shake down for the copy-cats. If you go into a competition expecting a fair leaderboard, you’re gonna be disappointed when public notebook forkers fill up the silver/bronze tiers at the end. The reality is the leaderboard (outside of the top few genuinely innovative gold notebooks) is mostly just based on who forked the highest score kernel fastest and who got lucky with a random seed. That’s just a fact at this point.\n\nAlso, many gold solutions at the top don't come from a standard 16gb consumer gpu. When the write ups come out, you will see massive ensembles that required burning hundreds of gpu hours (I am not denying building at huge scale also requires an expertise, having access to hardware capacity isn't equal to win ofcourse but if we are talking about fairness lets put it down). They are not usually something a typical home user can do without spending a fortune on cloud costs. If we’re lucky, maybe 1 or 2 real experts actually beat the game with a clever approach that fits on a normal scale, and we get to read and learn from them. Unless you are real expert in field so that you can beat the scale with innovation, IMO the only reason to even attend a competition now is for the learning opportunity and getting your hands dirty on a new subject. Expecting a fair rank for hard work is just asking for a headache. Sorry to hear that you spent alot of time doing the boring data extraction & cleaning staff, lets hope there will be a shakedown and genuinely hard work like yours gets rewarded this time. 🍀", "votes": 6 }, { "id": 3410584, "postDate": "2026-02-23T10:17:28.823Z", "content": "i see but i need to keep going as if i leave now whole progress will be lost , so lets hope for a medal and good rank
", "rawMarkdown": "i see but i need to keep going as if i leave now whole progress will be lost , so lets hope for a medal and good rank", "votes": 2 } ] } ] }, { "id": 3408437, "postDate": "2026-02-20T16:09:13.917Z", "content": "In the public leader board i can see around 350 people stuck at 35.1 score. I have trained my model with 77k rows high quality self made dataset, written massive codes to extract data from different pdf , book , website etc which took me 2.5 weeks still i am at 34 idk is there something i am missing or 800 people above are also collecting that much data like akkadian data is not easy to find and extract
\nif you see the board under 34 score at increase of 1 point you will mostly cross 40-50 teams but here between 34 and 35 you are crossing 700 teamssss , i dont have any words The score is accommodated between 34 and 35 how is that possible???????/
\nplease someone tell me what is happening up there
", "rawMarkdown": "In the public leader board i can see around 350 people stuck at 35.1 score. I have trained my model with 77k rows high quality self made dataset, written massive codes to extract data from different pdf , book , website etc which took me 2.5 weeks still i am at 34 idk is there something i am missing or 800 people above are also collecting that much data like akkadian data is not easy to find and extract\n\nif you see the board under 34 score at increase of 1 point you will mostly cross 40-50 teams but here between 34 and 35 you are crossing 700 teamssss , i dont have any words The score is accommodated between 34 and 35 how is that possible???????/\n\nplease someone tell me what is happening up there", "votes": 6 }, { "id": 3408498, "postDate": "2026-02-20T18:59:07.540Z", "content": "that score is pre-update, if you submit now it is after-update score, which means your 34 is a solid score. once the rescoring is underway you will likely see those public score got dropped by a lot probably.
", "rawMarkdown": "that score is pre-update, if you submit now it is after-update score, which means your 34 is a solid score. once the rescoring is underway you will likely see those public score got dropped by a lot probably.", "votes": 2, "replies": [ { "id": 3408612, "postDate": "2026-02-21T01:00:21.780Z", "content": "I see so they have over fitted on test set right?
", "rawMarkdown": "I see so they have over fitted on test set right?", "replies": [ { "id": 3408626, "postDate": "2026-02-21T02:02:39.707Z", "content": "the previous test set yes, not the new one
", "rawMarkdown": "the previous test set yes, not the new one", "votes": 1 } ] }, { "id": 3408675, "postDate": "2026-02-21T05:04:29.260Z", "content": "\n\nif you submit now it is after-update score
\n
don't think so, there is no re-scoring yet
", "rawMarkdown": "> if you submit now it is after-update score\n\ndon't think so, there is no re-scoring yet", "replies": [ { "id": 3408682, "postDate": "2026-02-21T05:35:16.760Z", "content": "no re-scoring on old submission, but if you submit now it will be scored on the new test
", "rawMarkdown": "no re-scoring on old submission, but if you submit now it will be scored on the new test", "votes": 2 }, { "id": 3408749, "postDate": "2026-02-21T08:40:33.127Z", "content": "When does that new test set got updated i think its same from 2-3 weeks
", "rawMarkdown": "When does that new test set got updated i think its same from 2-3 weeks\n" }, { "id": 3408759, "postDate": "2026-02-21T09:20:09.880Z", "content": "It probably did update. I found out yesterday that I couldn't reproduce my best score and spent ages debugging it. Reading this thread today, everything finally makes sense!
", "rawMarkdown": "It probably did update. I found out yesterday that I couldn't reproduce my best score and spent ages debugging it. Reading this thread today, everything finally makes sense!", "votes": 1 } ] } ] }, { "id": 3413203, "postDate": "2026-02-24T13:15:44.790Z", "content": "might be an issue of dadaset? @adarsh2626
", "rawMarkdown": "might be an issue of dadaset? @adarsh2626 ", "replies": [ { "id": 3413230, "postDate": "2026-02-24T14:11:02.810Z", "content": "Na there is not\nI saw that model that were getting 35.2 or 35.4 in new test set are getting 33.1 in some ones notebook without any code change but mine in getting 34.9 on new set
", "rawMarkdown": "Na there is not\nI saw that model that were getting 35.2 or 35.4 in new test set are getting 33.1 in some ones notebook without any code change but mine in getting 34.9 on new set " } ] }, { "id": 3409056, "postDate": "2026-02-22T01:58:37.307Z", "content": "Regarding your high-quality custom dataset of 77,000 rows—what data sources are you using besides the provided train.csv? Is this data provided by the organizers, or is it something you collected yourself?
\nI am currently preprocessing the train.csv and working on data augmentation, but even after splitting the roughly 1,500 rows, I am nowhere near reaching 77,000 entries.
", "rawMarkdown": "Regarding your high-quality custom dataset of 77,000 rows—what data sources are you using besides the provided train.csv? Is this data provided by the organizers, or is it something you collected yourself?\n\nI am currently preprocessing the train.csv and working on data augmentation, but even after splitting the roughly 1,500 rows, I am nowhere near reaching 77,000 entries.", "replies": [ { "id": 3409066, "postDate": "2026-02-22T03:01:50.347Z", "content": "I have made my self , have written 1000s of line of code to extract and clean the code like when you extrat (s) is copied hai # and 1/3 as * and many more so have to deal with that by understanding the data and changing the code according to it
\nIf you want to know about the source of data contact me on linkdin from my kaggle profile
", "rawMarkdown": "I have made my self , have written 1000s of line of code to extract and clean the code like when you extrat (s) is copied hai # and 1/3 as * and many more so have to deal with that by understanding the data and changing the code according to it\n\nIf you want to know about the source of data contact me on linkdin from my kaggle profile ", "replies": [ { "id": 3409140, "postDate": "2026-02-22T07:59:19.417Z", "content": "What's the source of the transliterations or did you made synthesized pairs..??
", "rawMarkdown": "What's the source of the transliterations or did you made synthesized pairs..??" }, { "id": 3409141, "postDate": "2026-02-22T08:02:36.203Z", "content": "its not a single source i collected many book and pdf understand them written different script for each it took me very long time to find sources and write the script
", "rawMarkdown": "its not a single source i collected many book and pdf understand them written different script for each it took me very long time to find sources and write the script" }, { "id": 3409148, "postDate": "2026-02-22T08:08:44.323Z", "content": "@adarsh2626 Why can't you just share the notebook here to get the data? Why making it closed source?
", "rawMarkdown": "@adarsh2626 Why can't you just share the notebook here to get the data? Why making it closed source?", "votes": -3 }, { "id": 3409150, "postDate": "2026-02-22T08:11:11.513Z", "content": "Does that improves the score like i used train.csv itself after segmentation i only got 34.2 as LB score but I tried to add more data but It only decreases the accuracy that's the issue here
", "rawMarkdown": "Does that improves the score like i used train.csv itself after segmentation i only got 34.2 as LB score but I tried to add more data but It only decreases the accuracy that's the issue here" }, { "id": 3409151, "postDate": "2026-02-22T08:12:46.397Z", "content": "@swami93 well its a competition and every one here is to win so sharing it will like to axe one's own foot every team up in board is doing same, i will do it after the competition ends.
\nPlease do not take it in bad sense
", "rawMarkdown": "@swami93 well its a competition and every one here is to win so sharing it will like to axe one's own foot every team up in board is doing same, i will do it after the competition ends.\n\nPlease do not take it in bad sense", "votes": 3 }, { "id": 3409155, "postDate": "2026-02-22T08:27:18.767Z", "content": "@antonyjoseph2121 using multiple-accounts is a quick way to get banned
", "rawMarkdown": "@antonyjoseph2121 using multiple-accounts is a quick way to get banned" }, { "id": 3409158, "postDate": "2026-02-22T08:33:40.350Z", "content": "This account is only for training i didn't give a single submission using this.I know use multiple accounts is against the rule 😅
", "rawMarkdown": "This account is only for training i didn't give a single submission using this.I know use multiple accounts is against the rule 😅" }, { "id": 3409164, "postDate": "2026-02-22T08:59:54.213Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 3409147, "postDate": "2026-02-22T08:08:18.677Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 3408489, "author_name": "moth", "author_url": "", "post_date": "2026-02-20T18:18:17.640000", "content": "All those people are just unoriginal forkers, people who are not truly interested in competing and rather copy/paste other people's best scoring public notebooks and submit them. That is why you see so many people with the same score. Don't worry, those public notebooks almost always are heavily overfitting the public leaderboard so in the end they get all shaked down.
", "votes": 7, "replies": [ { "id": 3408543, "author_name": "Aaron Bornstein", "author_url": "", "post_date": "2026-02-20T20:03:46.960000", "content": "exactly this
", "votes": 0, "replies": [] }, { "id": 3408613, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-21T01:01:14.603000", "content": "Feeling good after hearing that
\nBut there should be some moderation over what is submitted like the people who are genuinely doing the hard work are betten buy someone who is just copy and pasting model no real work
", "votes": 1, "replies": [ { "id": 3410547, "author_name": "Pelin Keskin", "author_url": "", "post_date": "2026-02-23T08:52:35.053000", "content": "This post shows how grim the situation is right now (https://www.kaggle.com/discussions/product-feedback/671127#3400122) and it feels like Kaggle doesn't really care… moth said public LB toppers \"almost always\" overfit, but honestly, that hasn't happened in most of the comps I've joined lately.
\nThere's a lot of medal corruption going on. Even some gold medal write-ups basically say \"oh I admit just merged the highest scoring public model in to my own and got lucky lol.\" From my experience, out of 5 competitions in the last year, only one actually had a real shake down for the copy-cats. If you go into a competition expecting a fair leaderboard, you’re gonna be disappointed when public notebook forkers fill up the silver/bronze tiers at the end. The reality is the leaderboard (outside of the top few genuinely innovative gold notebooks) is mostly just based on who forked the highest score kernel fastest and who got lucky with a random seed. That’s just a fact at this point.
\nAlso, many gold solutions at the top don't come from a standard 16gb consumer gpu. When the write ups come out, you will see massive ensembles that required burning hundreds of gpu hours (I am not denying building at huge scale also requires an expertise, having access to hardware capacity isn't equal to win ofcourse but if we are talking about fairness lets put it down). They are not usually something a typical home user can do without spending a fortune on cloud costs. If we’re lucky, maybe 1 or 2 real experts actually beat the game with a clever approach that fits on a normal scale, and we get to read and learn from them. Unless you are real expert in field so that you can beat the scale with innovation, IMO the only reason to even attend a competition now is for the learning opportunity and getting your hands dirty on a new subject. Expecting a fair rank for hard work is just asking for a headache. Sorry to hear that you spent alot of time doing the boring data extraction & cleaning staff, lets hope there will be a shakedown and genuinely hard work like yours gets rewarded this time. 🍀
", "votes": 6, "replies": [] }, { "id": 3410584, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-23T10:17:28.823000", "content": "i see but i need to keep going as if i leave now whole progress will be lost , so lets hope for a medal and good rank
", "votes": 2, "replies": [] } ] } ] }, { "id": 3408498, "author_name": "hongan", "author_url": "", "post_date": "2026-02-20T18:59:07.540000", "content": "that score is pre-update, if you submit now it is after-update score, which means your 34 is a solid score. once the rescoring is underway you will likely see those public score got dropped by a lot probably.
", "votes": 2, "replies": [ { "id": 3408612, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-21T01:00:21.780000", "content": "I see so they have over fitted on test set right?
", "votes": 0, "replies": [ { "id": 3408626, "author_name": "hongan", "author_url": "", "post_date": "2026-02-21T02:02:39.707000", "content": "the previous test set yes, not the new one
", "votes": 1, "replies": [] } ] }, { "id": 3408675, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2026-02-21T05:04:29.260000", "content": "\n\nif you submit now it is after-update score
\n
don't think so, there is no re-scoring yet
", "votes": 0, "replies": [ { "id": 3408682, "author_name": "hongan", "author_url": "", "post_date": "2026-02-21T05:35:16.760000", "content": "no re-scoring on old submission, but if you submit now it will be scored on the new test
", "votes": 2, "replies": [] }, { "id": 3408749, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-21T08:40:33.127000", "content": "When does that new test set got updated i think its same from 2-3 weeks
", "votes": 0, "replies": [] }, { "id": 3408759, "author_name": "To be happy", "author_url": "", "post_date": "2026-02-21T09:20:09.880000", "content": "It probably did update. I found out yesterday that I couldn't reproduce my best score and spent ages debugging it. Reading this thread today, everything finally makes sense!
", "votes": 1, "replies": [] } ] } ] }, { "id": 3413203, "author_name": "Navneet", "author_url": "", "post_date": "2026-02-24T13:15:44.790000", "content": "might be an issue of dadaset? @adarsh2626
", "votes": 0, "replies": [ { "id": 3413230, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-24T14:11:02.810000", "content": "Na there is not\nI saw that model that were getting 35.2 or 35.4 in new test set are getting 33.1 in some ones notebook without any code change but mine in getting 34.9 on new set
", "votes": 0, "replies": [] } ] }, { "id": 3409056, "author_name": "YShunT", "author_url": "", "post_date": "2026-02-22T01:58:37.307000", "content": "Regarding your high-quality custom dataset of 77,000 rows—what data sources are you using besides the provided train.csv? Is this data provided by the organizers, or is it something you collected yourself?
\nI am currently preprocessing the train.csv and working on data augmentation, but even after splitting the roughly 1,500 rows, I am nowhere near reaching 77,000 entries.
", "votes": 0, "replies": [ { "id": 3409066, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-22T03:01:50.347000", "content": "I have made my self , have written 1000s of line of code to extract and clean the code like when you extrat (s) is copied hai # and 1/3 as * and many more so have to deal with that by understanding the data and changing the code according to it
\nIf you want to know about the source of data contact me on linkdin from my kaggle profile
", "votes": 0, "replies": [ { "id": 3409140, "author_name": "Antony Joseph2121", "author_url": "", "post_date": "2026-02-22T07:59:19.417000", "content": "What's the source of the transliterations or did you made synthesized pairs..??
", "votes": 0, "replies": [] }, { "id": 3409141, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-22T08:02:36.203000", "content": "its not a single source i collected many book and pdf understand them written different script for each it took me very long time to find sources and write the script
", "votes": 0, "replies": [] }, { "id": 3409148, "author_name": "Swami93", "author_url": "", "post_date": "2026-02-22T08:08:44.323000", "content": "@adarsh2626 Why can't you just share the notebook here to get the data? Why making it closed source?
", "votes": -3, "replies": [] }, { "id": 3409150, "author_name": "Antony Joseph2121", "author_url": "", "post_date": "2026-02-22T08:11:11.513000", "content": "Does that improves the score like i used train.csv itself after segmentation i only got 34.2 as LB score but I tried to add more data but It only decreases the accuracy that's the issue here
", "votes": 0, "replies": [] }, { "id": 3409151, "author_name": "Adarsh Kumar", "author_url": "", "post_date": "2026-02-22T08:12:46.397000", "content": "@swami93 well its a competition and every one here is to win so sharing it will like to axe one's own foot every team up in board is doing same, i will do it after the competition ends.
\nPlease do not take it in bad sense
", "votes": 3, "replies": [] }, { "id": 3409155, "author_name": "Yurnero", "author_url": "", "post_date": "2026-02-22T08:27:18.767000", "content": "@antonyjoseph2121 using multiple-accounts is a quick way to get banned
", "votes": 0, "replies": [] }, { "id": 3409158, "author_name": "Antony Joseph2121", "author_url": "", "post_date": "2026-02-22T08:33:40.350000", "content": "This account is only for training i didn't give a single submission using this.I know use multiple accounts is against the rule 😅
", "votes": 0, "replies": [] }, { "id": 3409164, "author_name": "", "author_url": "", "post_date": "2026-02-22T08:59:54.213000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 3409147, "author_name": "", "author_url": "", "post_date": "2026-02-22T08:08:18.677000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3408489": "All those people are just unoriginal forkers, people who are not truly interested in competing and rather copy/paste other people's best scoring public notebooks and submit them. That is why you see so many people with the same score. Don't worry, those public notebooks almost always are heavily overfitting the public leaderboard so in the end they get all shaked down.", "3408437": "In the public leader board i can see around 350 people stuck at 35.1 score. I have trained my model with 77k rows high quality self made dataset, written massive codes to extract data from different pdf , book , website etc which took me 2.5 weeks still i am at 34 idk is there something i am missing or 800 people above are also collecting that much data like akkadian data is not easy to find and extract\n\nif you see the board under 34 score at increase of 1 point you will mostly cross 40-50 teams but here between 34 and 35 you are crossing 700 teamssss , i dont have any words The score is accommodated between 34 and 35 how is that possible???????/\n\nplease someone tell me what is happening up there", "3408498": "that score is pre-update, if you submit now it is after-update score, which means your 34 is a solid score. once the rescoring is underway you will likely see those public score got dropped by a lot probably.", "3413203": "might be an issue of dadaset? @adarsh2626 ", "3409056": "Regarding your high-quality custom dataset of 77,000 rows—what data sources are you using besides the provided train.csv? Is this data provided by the organizers, or is it something you collected yourself?\n\nI am currently preprocessing the train.csv and working on data augmentation, but even after splitting the roughly 1,500 rows, I am nowhere near reaching 77,000 entries." } }