Someone takes an existing top notebook, changes N_SPLITS from 5 to 10 or weights, reruns it, gets +0.01 LB boost, and publishes it.
\nThat clone gets 100+ upvotes or even a gold medal. People who’ve done no actual work in the competition just run these notebooks and end up high on the leaderboard.
\nWhat’s the point of this system? I’m new here, just trying to understand.
", "messages": [ { "id": 3249312, "postDate": "2025-07-16T08:07:04.943Z", "content": "Someone takes an existing top notebook, changes N_SPLITS from 5 to 10 or weights, reruns it, gets +0.01 LB boost, and publishes it.
\nThat clone gets 100+ upvotes or even a gold medal. People who’ve done no actual work in the competition just run these notebooks and end up high on the leaderboard.
\nWhat’s the point of this system? I’m new here, just trying to understand.
", "rawMarkdown": "Someone takes an existing top notebook, changes N_SPLITS from 5 to 10 or weights, reruns it, gets +0.01 LB boost, and publishes it.\n\nThat clone gets 100+ upvotes or even a gold medal. People who’ve done no actual work in the competition just run these notebooks and end up high on the leaderboard.\n\nWhat’s the point of this system? I’m new here, just trying to understand.", "votes": 18 }, { "id": 3249962, "postDate": "2025-07-17T12:50:37.987Z", "content": "Check out the private leaderboard from a previous competition by the same host to see how much it can differ from the public leaderboard: https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard
\nIn that competition, the 1st place solution increased by 1,488 ranks, and the 2nd place increased by 2,603. Many Kaggle competitions end with a significant public LB shake-up like this.
Then there are some competitions where free loaders end up with medals. A recent example is the March Machine Learning Mania 2025 competition: https://www.kaggle.com/competitions/march-machine-learning-mania-2025
\nIn that one, 69 notebooks had the exact same LB score, most likely from copying the same 2 public notebooks. But here's the thing: over 3,000 data scientists followed that competition, and many probably noticed. So even if someone got a silver, it came with doubt and reputation risk…
Bottom line: take the public leaderboard with a grain of salt. Avoid using public approaches you don’t fully understand or trust. And most importantly, rely on your own testing scheme rather than chasing small public LB gains. IMO most importantly kaggle is a platform for practice and learning and medals are nice little encouragements. If you aren't working it out yourself, you aren't improving your expertise, then what is the point of participating anyway.
", "rawMarkdown": "Check out the private leaderboard from a previous competition by the same host to see how much it can differ from the public leaderboard: https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard\nIn that competition, the 1st place solution increased by 1,488 ranks, and the 2nd place increased by 2,603. Many Kaggle competitions end with a significant public LB shake-up like this.\n\nThen there are some competitions where free loaders end up with medals. A recent example is the March Machine Learning Mania 2025 competition: https://www.kaggle.com/competitions/march-machine-learning-mania-2025 \nIn that one, 69 notebooks had the exact same LB score, most likely from copying the same 2 public notebooks. But here's the thing: over 3,000 data scientists followed that competition, and many probably noticed. So even if someone got a silver, it came with doubt and reputation risk...\n\nBottom line: take the public leaderboard with a grain of salt. Avoid using public approaches you don’t fully understand or trust. And most importantly, rely on your own testing scheme rather than chasing small public LB gains. IMO most importantly kaggle is a platform for practice and learning and medals are nice little encouragements. If you aren't working it out yourself, you aren't improving your expertise, then what is the point of participating anyway. ", "votes": 6, "replies": [ { "id": 3249975, "postDate": "2025-07-17T13:11:25.643Z", "rawMarkdown": "", "votes": -2, "isDeleted": true, "replies": [ { "id": 3250046, "postDate": "2025-07-17T15:29:43.300Z", "content": "In this competition, the public leaderboard is based on just 42% of the test set — so the remaining 58% used for the private LB can cause a shake-up. I haven’t really looked closely at the top-scoring notebooks in this one, so I can’t say whether they’re solid, leak-free solutions with stable CV scores, or if we’ll see free loaders get shuffled down when the private LB is revealed. Time will tell…
", "rawMarkdown": "In this competition, the public leaderboard is based on just 42% of the test set — so the remaining 58% used for the private LB can cause a shake-up. I haven’t really looked closely at the top-scoring notebooks in this one, so I can’t say whether they’re solid, leak-free solutions with stable CV scores, or if we’ll see free loaders get shuffled down when the private LB is revealed. Time will tell...", "votes": -1 }, { "id": 3250065, "postDate": "2025-07-17T16:18:20.987Z", "rawMarkdown": "", "votes": -1, "isDeleted": true } ] }, { "id": 3250179, "postDate": "2025-07-17T20:30:21.297Z", "content": "I agree with this 100%.
\nbetter to lose the medal and gain the know-how than the other way around.
", "rawMarkdown": "I agree with this 100%. \n\nbetter to lose the medal and gain the know-how than the other way around. ", "votes": 2 } ] }, { "id": 3249398, "postDate": "2025-07-16T12:03:02.560Z", "content": "But this way you cannot win gold, extremely merge and tune public notebook just can let you win silver and no one will care the final solution you write. And most important thing, you do not learn anything in this competition. How you discover the relationship between these physical measurements and the gestures is what you need to learn and gain experience.
", "rawMarkdown": "But this way you cannot win gold, extremely merge and tune public notebook just can let you win silver and no one will care the final solution you write. And most important thing, you do not learn anything in this competition. How you discover the relationship between these physical measurements and the gestures is what you need to learn and gain experience.\n", "votes": 6, "replies": [ { "id": 3249920, "postDate": "2025-07-17T11:09:17.953Z", "content": "I completely agree with you. I’ve reproduced many winning solutions, but they often only provide the final answer (though I truly appreciate them). This approach only helps me understand how to solve past competitions. If you could also share your experimental results and step-by-step analysis, it would be much more informative for beginner Kagglers like me.
", "rawMarkdown": "I completely agree with you. I’ve reproduced many winning solutions, but they often only provide the final answer (though I truly appreciate them). This approach only helps me understand how to solve past competitions. If you could also share your experimental results and step-by-step analysis, it would be much more informative for beginner Kagglers like me." } ] }, { "id": 3251098, "postDate": "2025-07-20T00:10:37.453Z", "content": "Totally get the frustration; here’s what’s really happening:
\nWhy tiny tweaks + reposts get traction
\nDoes this hurt the competition?\nNot really. Public LB ≠ final ranking. Shallow clones rarely hold once the private LB (or second stage) scores drop. Serious competitors rely on local CV, not public noise.
\nHow to spot real contribution
\nWhy clones still appear
\nWhat to do instead of worrying
\nImproving your own signal
\nSanity checks before praising a “+0.01” notebook
\nMindset\nTreat public LB improvements under ~0.02 as “unconfirmed” until replicated in your CV. Your edge comes from rigor, not chasing every micro‑gain someone posts.
\nIn short: the system surfaces noise, but it doesn’t decide the final outcome. Build real skill: solid validation + genuine innovation. Let the clones farm early upvotes; they almost never win medals that matter (final private LB).
\nWelcome aboard—focus on craft, you’ll rise past the noise.
", "rawMarkdown": "Totally get the frustration; here’s what’s really happening:\n\n**Why tiny tweaks + reposts get traction**\n\n* Early visibility + recency bias: a freshly published notebook with a small LB bump surfaces in “Recently Published” and people auto‑upvote.\n* Many newcomers want a ready notebook to fork; they reward anything that *runs* and shows +0.01, even if it’s just noise.\n* A 0.01 public LB gain often *is* variance (different seed / KFold split) rather than real improvement.\n\n**Does this hurt the competition?**\nNot really. Public LB ≠ final ranking. Shallow clones rarely hold once the private LB (or second stage) scores drop. Serious competitors rely on local CV, not public noise.\n\n**How to spot real contribution**\n\n* Clear, reasoned CV strategy (stratification rationale, leak checks, stability stats).\n* Novel features / preprocessing / architecture changes explained and *validated* (tables of fold scores, seed averages).\n* Error analysis, robustness checks, inference speedups, ablation studies.\n* Reproducibility: fixed seeds, version pinning, concise modular code.\n\n**Why clones still appear**\n\n* Low barrier to entry encourages learning by modification (not evil, just low signal).\n* Medals reflect aggregate community voting; the system can’t perfectly judge originality.\n\n**What to do instead of worrying**\n\n* Ignore superficial forks; focus on building a rock‑solid offline CV that predicts *both* public + private.\n* Track your seeds; average scores; log fold dispersion.\n* Run ablations: prove each claimed improvement.\n* Share *one* well‑documented notebook late enough (or with hidden tricks locally) if you care about protecting ideas.\n* Use Discussions for deeper write‑ups (methodology > code dump). Those attract informed voters.\n\n**Improving your own signal**\n\n* Publish: (1) Clean baseline with robust CV; (2) Feature / model expansion with ablations; (3) Final inference / ensemble write‑up.\n* Include: dataset issues found, leakage tests, outlier handling, model diagnostics.\n* Time your publication when noise is lower (not right after launch surge).\n\n**Sanity checks before praising a “+0.01” notebook**\n\n* Compare its CV vs its claimed LB jump.\n* Look for seed averaging; absence often means variance exploitation.\n* Diff the code: are changes substantive or just N\\_SPLITS, learning rate, seed?\n\n**Mindset**\nTreat public LB improvements under \\~0.02 as “unconfirmed” until replicated in your CV. Your edge comes from rigor, not chasing every micro‑gain someone posts.\n\nIn short: the system surfaces noise, but it doesn’t decide the final outcome. Build real skill: solid validation + genuine innovation. Let the clones farm early upvotes; they almost never win medals that matter (final private LB).\n\nWelcome aboard—focus on craft, you’ll rise past the noise.", "votes": 3, "replies": [ { "id": 3251183, "postDate": "2025-07-20T06:13:30.557Z", "content": "I carefully read everything and truly appreciate you sharing such valuable insights.
", "rawMarkdown": "I carefully read everything and truly appreciate you sharing such valuable insights.", "votes": 1 }, { "id": 3251222, "postDate": "2025-07-20T08:13:03.193Z", "content": "Thank you so much for the insight <3
", "rawMarkdown": "Thank you so much for the insight <3\n", "isDeleted": true } ] }, { "id": 3252004, "postDate": "2025-07-21T21:32:24.837Z", "content": "I totally feel you! it’s frustrating to see low-effort clones rocket up the LB while original work gets buried.
\nFrom my experience, there is a messy reality that covers the good , the bad and the ugly.
\nThe Good! Tiny tweaks can legitimately improve models (say more folds = better CV = smarter stacking). Sometimes that +0.01 is the difference between 1st and 10th.
\nThen there is the Bad. Most of these are just noise. People cargo-culting changes without understanding why they work (or don’t). Upvotes often mean “this gave me free LB points,” not “this is insightful.”
\nAnd then finally the Ugly. Kaggle’s reward system does incentivize this. Medals get handed out for incremental gains, even if the core idea isn’t yours.
\nBut I think the secret is that the real leaderboard is the skills you build. The folks grinding on novel features, debugging leaks, or dissecting winners’ code. They’re the ones who’ll dominate the next comp. Clone wars might win short-term karma, but they don’t teach you anything.
\nStill sucks though. I wish originality carried more weight in medals.
", "rawMarkdown": "I totally feel you! it’s frustrating to see low-effort clones rocket up the LB while original work gets buried. \n\nFrom my experience, there is a messy reality that covers the good , the bad and the ugly. \n\nThe Good! Tiny tweaks can legitimately improve models (say more folds = better CV = smarter stacking). Sometimes that +0.01 is the difference between 1st and 10th.\n\nThen there is the Bad. Most of these are just noise. People cargo-culting changes without understanding why they work (or don’t). Upvotes often mean “this gave me free LB points,” not “this is insightful.”\n\nAnd then finally the Ugly. Kaggle’s reward system does incentivize this. Medals get handed out for incremental gains, even if the core idea isn’t yours.\n\nBut I think the secret is that the real leaderboard is the skills you build. The folks grinding on novel features, debugging leaks, or dissecting winners’ code. They’re the ones who’ll dominate the next comp. Clone wars might win short-term karma, but they don’t teach you anything.\n\nStill sucks though. I wish originality carried more weight in medals.", "votes": 3 }, { "id": 3249384, "postDate": "2025-07-16T11:13:22.217Z", "content": "I can only speak for competition medals. IMO, it's harder to achieve than you'd imagine. Usually, you'll need to at the very least beat the public baseline, and with something different, or novel, enough to differentiate.
\nLB standing now is not really meaningful (well, excluding the gold zone obviously), since we have about 2 months to go. My observation is that the last 2 weeks are more crucial, as it's the window when GMs join, and when teams merge.
\nIn any case, i'll offer you a positive bit for high scoring public notebooks:
\nHigher scoring public notebooks -> more NPCs enter to fork and submit -> more spots for medals (since the number of spots is proportional to teams which made at least one submission)
1854 teams at time of writing, 0.84 notebook released 8 hours ago. you can probably download the LB data now, and check in 1-2 days to see the number of 0.84 LB with less than 10 subs maybe to see how much the notebook attracted new teams.
", "rawMarkdown": "I can only speak for competition medals. IMO, it's harder to achieve than you'd imagine. Usually, you'll need to at the very least beat the public baseline, and with something different, or novel, enough to differentiate. \n\nLB standing now is not really meaningful (well, excluding the gold zone obviously), since we have about 2 months to go. My observation is that the last 2 weeks are more crucial, as it's the window when GMs join, and when teams merge.\n\nIn any case, i'll offer you a positive bit for high scoring public notebooks: \nHigher scoring public notebooks -> more NPCs enter to fork and submit -> more spots for medals (since the number of spots is proportional to teams which made at least one submission)\n\n1854 teams at time of writing, 0.84 notebook released 8 hours ago. you can probably download the LB data now, and check in 1-2 days to see the number of 0.84 LB with less than 10 subs maybe to see how much the notebook attracted new teams.", "votes": 3, "replies": [ { "id": 3249388, "postDate": "2025-07-16T11:25:12.247Z", "content": "Thanks for the detailed and helpful explanation — I appreciate it.
\nOne thing I’m still trying to understand is the actual value or impact of notebook medals.
\nBecause from what I’ve seen, some people make very minor changes to existing notebooks and still get gold medals and a lot of upvotes — even if they’ve done no real work themselves.
So I’m wondering, how meaningful are these medals?
", "rawMarkdown": "Thanks for the detailed and helpful explanation — I appreciate it.\n\nOne thing I’m still trying to understand is the actual value or impact of notebook medals.\nBecause from what I’ve seen, some people make very minor changes to existing notebooks and still get gold medals and a lot of upvotes — even if they’ve done no real work themselves.\n\nSo I’m wondering, how meaningful are these medals?", "votes": -1, "replies": [ { "id": 3249397, "postDate": "2025-07-16T12:00:04.137Z", "content": "same viewpoint as @shlomoron - thats how kaggle work for competition notebooks at least: i am quite sure majority are just doing a sort by Public Score, so higher score notebook usually get more coverage in that sense.
Also, the recent update (i believe), gives upvote for every fork, so it's no surprise that high scores notebook have high correlation with upvote (more coverage + more forks).
", "rawMarkdown": "same viewpoint as @shlomoron - thats how kaggle work for competition notebooks at least: i am quite sure majority are just doing a sort by `Public Score`, so higher score notebook usually get more coverage in that sense.\n\nAlso, the recent update (i believe), gives upvote for every fork, so it's no surprise that high scores notebook have high correlation with upvote (more coverage + more forks). ", "votes": 1 } ] } ] }, { "id": 3249409, "postDate": "2025-07-16T12:20:14.833Z", "content": "Public notebooks are just a helpful resource. Sure, people love seeing themselves in the top 200-300 on Kaggle, but real progress comes from getting inspired and grinding it out. I’d say my result is 50% vibe coding, 45% public notebooks, and 5% brainpower… well, maybe even less 😅
", "rawMarkdown": "Public notebooks are just a helpful resource. Sure, people love seeing themselves in the top 200-300 on Kaggle, but real progress comes from getting inspired and grinding it out. I’d say my result is 50% vibe coding, 45% public notebooks, and 5% brainpower... well, maybe even less 😅", "votes": 1, "replies": [ { "id": 3249612, "postDate": "2025-07-16T18:57:18.030Z", "content": "Haha, I’ve been hearing a lot about vibe coding recently. What have you found works best for data science applications in terms of models and workflows? Just curious what high level people have found to work best! -Thanks
", "rawMarkdown": "Haha, I’ve been hearing a lot about vibe coding recently. What have you found works best for data science applications in terms of models and workflows? Just curious what high level people have found to work best! -Thanks", "replies": [ { "id": 3249630, "postDate": "2025-07-16T19:44:03.770Z", "content": "I use different LLMs, designing with one, critiquing with another, debugging, and training. If one doesn’t work, I switch to another LLM. I draw a lot of inspiration from public notebooks. There are some best practices online 🙂
", "rawMarkdown": "I use different LLMs, designing with one, critiquing with another, debugging, and training. If one doesn’t work, I switch to another LLM. I draw a lot of inspiration from public notebooks. There are some best practices online 🙂", "votes": 1 } ] } ] }, { "id": 3253279, "postDate": "2025-07-24T11:16:21.083Z", "content": "1- Trust you CV score.
\n2- LB is a good measure of how you are doing but its not final.
\n3- View leaderboard of last CMI competition \"Problematic Internet Use\" you will see that the people on top had jumped more than 1000 or 2000 points after final scoring.
\n4- Don't worry about what people are doing. Just give your best.
Good luck
", "rawMarkdown": "1- Trust you CV score.\n2- LB is a good measure of how you are doing but its not final.\n3- View leaderboard of last CMI competition \"Problematic Internet Use\" you will see that the people on top had jumped more than 1000 or 2000 points after final scoring.\n4- Don't worry about what people are doing. Just give your best.\n\nGood luck", "votes": 2 }, { "id": 3250938, "postDate": "2025-07-19T14:12:33.357Z", "content": "\nYes, they've basically made notebook medals without any value for the moment.
\nHow this will play out longer term:
\nUntil one of these things happens, notebook GM status is nonsense and is rapidly getting diluted. They can retroactively rug users again, but they don't like removing GM status and having a GM status without actually having it feels humiliating.
\nI've opted out of the progression tiers as they are pretty silly until it stabilizes or is fixed.
\nHonestly, I like the tweaked notebooks, read them all, and would publish some myself.
\nBut obv they are not gold medal worthy. And much higher value notebooks are being ignored because of it. If I didn't know better, I'd assume they are trying to destroy the noteboook progression tier.
", "rawMarkdown": "@mylesoneill \n\nYes, they've basically made notebook medals without any value for the moment. \n\nHow this will play out longer term:\n\n1. People will stop forking these tweaked forks (unlikely, and sort of irritating to have to avoid forking a notebook)\n2. Kaggle will warn people for making forks without huge improvements (it's actually against the guidelines, believe it or not, though I really don't know why. It's the basis of opensource and these small improvements when added up in totality are often quite useful!! The whole concept of opensource is leveraging the crowd for small contributions)\n3. Kaggle will increase # of notebook medals (really bad idea - as it punishes people who make thoughtful notebooks)\n4. Kaggle will increase the level for fork votes to master or gm level (my fav option)\n5. Kaggle will remove the notebook progession tier\n6. Kaggle will tweak it in some other way\n\nUntil one of these things happens, notebook GM status is nonsense and is rapidly getting diluted. They can retroactively rug users again, but they don't like removing GM status and having a GM status without actually having it feels humiliating.\n\nI've opted out of the progression tiers as they are pretty silly until it stabilizes or is fixed.\n\nHonestly, I like the tweaked notebooks, read them all, and would publish some myself. \n\nBut obv they are not gold medal worthy. And much higher value notebooks are being ignored because of it. If I didn't know better, I'd assume they are trying to destroy the noteboook progression tier.", "votes": 2, "replies": [ { "id": 3250944, "postDate": "2025-07-19T14:36:47.773Z", "content": "Absolutely agree, this captures the issue perfectly.
", "rawMarkdown": "Absolutely agree, this captures the issue perfectly." } ] }, { "id": 3249958, "postDate": "2025-07-17T12:44:20.207Z", "content": "It's fine to accept that, in a past competition it was happening the same and look at the final results CMI problematic internet use.
\nAll the copied kernel will fall in the same pit at evaluation, effectively eliminating tons of submissions.
", "rawMarkdown": "It's fine to accept that, in a past competition it was happening the same and look at the final results [CMI problematic internet use](https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard).\n\nAll the copied kernel will fall in the same pit at evaluation, effectively eliminating tons of submissions. \n\n", "votes": 2 }, { "id": 3252790, "postDate": "2025-07-23T12:15:24.423Z", "content": "This is a really important question—and one that many new (and even experienced) Kagglers ask at some point.
\nKaggle encourages sharing and collaboration, which is part of what makes it such a vibrant learning community. But yes, there's a fine line between building upon others’ work and cloning for easy gains. When a minor change to a high-performing notebook gets heavily upvoted or even rewarded with medals, it can feel unfair—especially to those investing serious time into original modeling or exploration.
", "rawMarkdown": "This is a really important question—and one that many new (and even experienced) Kagglers ask at some point.\n\nKaggle encourages sharing and collaboration, which is part of what makes it such a vibrant learning community. But yes, there's a fine line between building upon others’ work and cloning for easy gains. When a minor change to a high-performing notebook gets heavily upvoted or even rewarded with medals, it can feel unfair—especially to those investing serious time into original modeling or exploration.", "votes": 1 }, { "id": 3249336, "postDate": "2025-07-16T09:02:54.860Z", "content": "That's how Kaggle works.
\nYou may think it's stupid, but then try to think on the alternatives and whether they will incentivise Kagglers to share more or less of the helpful content.
I’m not against sharing. But getting a medal for changing one parameter in someone else’s notebook isn’t sharing.
\nI think this doesn’t encourage contribution — it encourages copy-paste.
", "rawMarkdown": "I’m not against sharing. But getting a medal for changing one parameter in someone else’s notebook isn’t sharing.\n\nI think this doesn’t encourage contribution — it encourages copy-paste. ", "votes": 1, "replies": [ { "id": 3249348, "postDate": "2025-07-16T09:35:08.457Z", "content": "So, I ask again, what is your alternative?
\nOnly, before you answer, please remember that Kagglers are very good at gaming any system.
I get that any system can be gamed, but it still feels worth questioning when very small changes lead to big rewards.
\nAlso, I’m just a user sharing an observation — coming up with alternatives isn’t really my role.
", "rawMarkdown": "I get that any system can be gamed, but it still feels worth questioning when very small changes lead to big rewards.\n\nAlso, I’m just a user sharing an observation — coming up with alternatives isn’t really my role.", "votes": -2 }, { "id": 3249366, "postDate": "2025-07-16T10:17:20.890Z", "content": "It's easy to criticize. But only criticism does not contribute much.
\nEspecially when it was heard so many times before.
\nYou have no idea how many times I have seen this observation already…I assure you that the Kaggle staff is aware.
\nUntil you or someone else comes up with a better alternative, this is the best we can have.
I understand, and I appreciate your perspective.
\nAs someone new here, I’m just trying to learn and make sense of how things work.
\nI think If this kind of feedback has been shared many times before, maybe it simply shows that some people feel strongly about it — which could be worth noting.
@shlomoron fork + submit of a public notebook without any changes or with exact same predictions should not be considered a valid submission and should not appear on the leaderboard.
\nThat is easy to implement for Kaggle, this does not penalize people who want to fork and change a few parameters, this still encourages people to share if they want to. But you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition). How about that proposal ?
", "rawMarkdown": "@shlomoron fork + submit of a public notebook without any changes or with exact same predictions should not be considered a valid submission and should not appear on the leaderboard.\n\nThat is easy to implement for Kaggle, this does not penalize people who want to fork and change a few parameters, this still encourages people to share if they want to. But you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition). How about that proposal ?", "votes": 1 }, { "id": 3249400, "postDate": "2025-07-16T12:07:11.700Z", "content": "I don't see the issue, winning a Kaggle comp is not about changing a few params of a public notebook. It is the ability to merge it with your current pipeline and gain an advantage.
", "rawMarkdown": "I don't see the issue, winning a Kaggle comp is not about changing a few params of a public notebook. It is the ability to merge it with your current pipeline and gain an advantage. ", "votes": 2 }, { "id": 3249425, "postDate": "2025-07-16T13:03:03.503Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3249428, "postDate": "2025-07-16T13:03:39.743Z", "content": "I really like the idea of excluding submissions with identical prediction files from the leaderboard — it’s a fair and effective solution.
\nThat said, a separate issue is users getting gold medals for copied notebooks with minimal effort (e.g., just changing a weight, N_SPLIT, or random_state).
Your proposal is quite useless, I will just add some extra lines of junk code so the code is not the same, and a very small noise to prediction (or will change the seed?).
\nIt adds complexity while solving nothing.
\n(And anyway it's not the issue we are talking about here- and if you ask me, it's not an issue at all).
\n\nBut you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition).
\n
There used to be a symbol \"< >\" on the LB to indicate which public notebook scored that on the LB (easier to ignore those repeated subs), though i havent seen it for a while so it may have been removed.
", "rawMarkdown": "> But you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition).\n\nThere used to be a symbol \"< >\" on the LB to indicate which public notebook scored that on the LB (easier to ignore those repeated subs), though i havent seen it for a while so it may have been removed.", "votes": 2 }, { "id": 3249621, "postDate": "2025-07-16T19:22:46.580Z", "content": "Did you see any gold medal solution write-up where someone says, \"I just changed the seed or splits and win a gold medal or silver\"?
\n@huseyingur, I haven't come across any like that.
\nyou have to work hard for gold, silver and winning.
There was. Place 5 in AIMO (the first one). He passed up the price after saying that he feel he does not deserve it.
\nBut it can happen only in conpetitions with high lottery factor.
This was a recent change. https://www.kaggle.com/discussions/product-announcements/588704
\nThey are probably hoping it will fix as the community self-policies and we are in a temporary gaming phase.
\nI'm 90% sure something will have to change as I mentioned above. Maybe 10 forks = 1 upvote or something. There are options.
\nI'm kinda hoping they will abandon the whole progression tier thing.
\nThe truth is the titles don't mean much to industry anymore as more specific skillsets, especially those which are LLM specific, are what is mostly sought after these days. Some comp wins are specific for certain verticals, but again, it's not the title. It's the comp you won.
\nI think removing the discussion tier really broke everything. They should have tried fixing it instead.
\nMaybe they should have got rid of the comp and notebook tier, perhaps and just kept discussions? Considering in discussions is where our colored circles actually show up, that makes more sense to me.
\nPeople who rank #1 in comps and don't participate much in the community, I be skeptical if they actually care about the titles.
", "rawMarkdown": "@huseyingur \n\nThis was a recent change. https://www.kaggle.com/discussions/product-announcements/588704\n\nThey are probably hoping it will fix as the community self-policies and we are in a temporary gaming phase.\n\nI'm 90% sure something will have to change as I mentioned above. Maybe 10 forks = 1 upvote or something. There are options.\n\nI'm kinda hoping they will abandon the whole progression tier thing. \n\nThe truth is the titles don't mean much to industry anymore as more specific skillsets, especially those which are LLM specific, are what is mostly sought after these days. Some comp wins are specific for certain verticals, but again, it's not the title. It's the comp you won.\n\nI think removing the discussion tier really broke everything. They should have tried fixing it instead. \n\nMaybe they should have got rid of the comp and notebook tier, perhaps and just kept discussions? Considering in discussions is where our colored circles actually show up, that makes more sense to me. \n\nPeople who rank #1 in comps and don't participate much in the community, I be skeptical if they actually care about the titles.\n\n\n", "votes": -1 }, { "id": 3251035, "postDate": "2025-07-19T18:49:20.087Z", "content": "Gold is non trivial, silver can be got by luck fairly easily. Consistently winning silver without piling up a lot of bronze can be harder.
\n\n\nDid you see any gold medal solution write-up where someone says, \"I just changed the seed or splits and win a gold medal or silver\"?
\n
\n @huseyingur, I haven't come across any like that.
\n you have to work hard for gold, silver and winning.
My first comp, I made a thoughtful, but not overly effortful, change to a notebook, and lucked out with a silver. In comps with poor cv / lb correlation that result in significant shakeups, this happens quite a bit. Luck plays a very big role in these comps, even for gold.
\nBut yes, consistently winning well is much harder for sure. You generally want to avoid bronze and even silver if you can.
", "rawMarkdown": "Gold is non trivial, silver can be got by luck fairly easily. Consistently winning silver without piling up a lot of bronze can be harder.\n\n> Did you see any gold medal solution write-up where someone says, \"I just changed the seed or splits and win a gold medal or silver\"?\n> @huseyingur, I haven't come across any like that. \n> you have to work hard for gold, silver and winning.\n\nMy first comp, I made a thoughtful, but not overly effortful, change to a notebook, and lucked out with a silver. In comps with poor cv / lb correlation that result in significant shakeups, this happens quite a bit. Luck plays a very big role in these comps, even for gold. \n\nBut yes, consistently winning well is much harder for sure. You generally want to avoid bronze and even silver if you can.", "votes": -1 } ] } ] }, { "id": 3249616, "postDate": "2025-07-16T19:12:35.670Z", "content": "I think sometimes luck might be on your side but just changing splits or seeds isn’t enough to win a silver or gold.
\nIn this competition specifically, most of the top public solutions are using saved models, so there's definitely more going on than just tuning seeds.
Well public notebooks are just a helpful resource if there is any doubt. Ofcourse, people do get distracted and opt an easier option by copying but real progress comes from getting inspired and grinding it out.
", "rawMarkdown": "Well public notebooks are just a helpful resource if there is any doubt. Ofcourse, people do get distracted and opt an easier option by copying but real progress comes from getting inspired and grinding it out." }, { "id": 3250849, "postDate": "2025-07-19T09:10:39.517Z", "content": "Public notebooks are just a helpful resource.
", "rawMarkdown": "Public notebooks are just a helpful resource." }, { "id": 3249368, "postDate": "2025-07-16T10:19:43.323Z", "rawMarkdown": "", "votes": -3, "isDeleted": true, "replies": [ { "id": 3249382, "postDate": "2025-07-16T11:09:16.387Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 3249962, "author_name": "Pelin Keskin", "author_url": "", "post_date": "2025-07-17T12:50:37.987000", "content": "Check out the private leaderboard from a previous competition by the same host to see how much it can differ from the public leaderboard: https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard
\nIn that competition, the 1st place solution increased by 1,488 ranks, and the 2nd place increased by 2,603. Many Kaggle competitions end with a significant public LB shake-up like this.
Then there are some competitions where free loaders end up with medals. A recent example is the March Machine Learning Mania 2025 competition: https://www.kaggle.com/competitions/march-machine-learning-mania-2025
\nIn that one, 69 notebooks had the exact same LB score, most likely from copying the same 2 public notebooks. But here's the thing: over 3,000 data scientists followed that competition, and many probably noticed. So even if someone got a silver, it came with doubt and reputation risk…
Bottom line: take the public leaderboard with a grain of salt. Avoid using public approaches you don’t fully understand or trust. And most importantly, rely on your own testing scheme rather than chasing small public LB gains. IMO most importantly kaggle is a platform for practice and learning and medals are nice little encouragements. If you aren't working it out yourself, you aren't improving your expertise, then what is the point of participating anyway.
", "votes": 6, "replies": [ { "id": 3249975, "author_name": "", "author_url": "", "post_date": "2025-07-17T13:11:25.643000", "content": "", "votes": -2, "replies": [ { "id": 3250046, "author_name": "Pelin Keskin", "author_url": "", "post_date": "2025-07-17T15:29:43.300000", "content": "In this competition, the public leaderboard is based on just 42% of the test set — so the remaining 58% used for the private LB can cause a shake-up. I haven’t really looked closely at the top-scoring notebooks in this one, so I can’t say whether they’re solid, leak-free solutions with stable CV scores, or if we’ll see free loaders get shuffled down when the private LB is revealed. Time will tell…
", "votes": -1, "replies": [] }, { "id": 3250065, "author_name": "", "author_url": "", "post_date": "2025-07-17T16:18:20.987000", "content": "", "votes": -1, "replies": [] } ] }, { "id": 3250179, "author_name": "Doron Fingold", "author_url": "", "post_date": "2025-07-17T20:30:21.297000", "content": "I agree with this 100%.
\nbetter to lose the medal and gain the know-how than the other way around.
", "votes": 2, "replies": [] } ] }, { "id": 3249398, "author_name": "Tom", "author_url": "", "post_date": "2025-07-16T12:03:02.560000", "content": "But this way you cannot win gold, extremely merge and tune public notebook just can let you win silver and no one will care the final solution you write. And most important thing, you do not learn anything in this competition. How you discover the relationship between these physical measurements and the gestures is what you need to learn and gain experience.
", "votes": 6, "replies": [ { "id": 3249920, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-07-17T11:09:17.953000", "content": "I completely agree with you. I’ve reproduced many winning solutions, but they often only provide the final answer (though I truly appreciate them). This approach only helps me understand how to solve past competitions. If you could also share your experimental results and step-by-step analysis, it would be much more informative for beginner Kagglers like me.
", "votes": 0, "replies": [] } ] }, { "id": 3251098, "author_name": "AnthonyTherrien", "author_url": "", "post_date": "2025-07-20T00:10:37.453000", "content": "Totally get the frustration; here’s what’s really happening:
\nWhy tiny tweaks + reposts get traction
\nDoes this hurt the competition?\nNot really. Public LB ≠ final ranking. Shallow clones rarely hold once the private LB (or second stage) scores drop. Serious competitors rely on local CV, not public noise.
\nHow to spot real contribution
\nWhy clones still appear
\nWhat to do instead of worrying
\nImproving your own signal
\nSanity checks before praising a “+0.01” notebook
\nMindset\nTreat public LB improvements under ~0.02 as “unconfirmed” until replicated in your CV. Your edge comes from rigor, not chasing every micro‑gain someone posts.
\nIn short: the system surfaces noise, but it doesn’t decide the final outcome. Build real skill: solid validation + genuine innovation. Let the clones farm early upvotes; they almost never win medals that matter (final private LB).
\nWelcome aboard—focus on craft, you’ll rise past the noise.
", "votes": 3, "replies": [ { "id": 3251183, "author_name": "Huseyin Gur", "author_url": "", "post_date": "2025-07-20T06:13:30.557000", "content": "I carefully read everything and truly appreciate you sharing such valuable insights.
", "votes": 1, "replies": [] }, { "id": 3251222, "author_name": "", "author_url": "", "post_date": "2025-07-20T08:13:03.193000", "content": "Thank you so much for the insight <3
", "votes": 0, "replies": [] } ] }, { "id": 3252004, "author_name": "Tamuno opubo Dappa", "author_url": "", "post_date": "2025-07-21T21:32:24.837000", "content": "I totally feel you! it’s frustrating to see low-effort clones rocket up the LB while original work gets buried.
\nFrom my experience, there is a messy reality that covers the good , the bad and the ugly.
\nThe Good! Tiny tweaks can legitimately improve models (say more folds = better CV = smarter stacking). Sometimes that +0.01 is the difference between 1st and 10th.
\nThen there is the Bad. Most of these are just noise. People cargo-culting changes without understanding why they work (or don’t). Upvotes often mean “this gave me free LB points,” not “this is insightful.”
\nAnd then finally the Ugly. Kaggle’s reward system does incentivize this. Medals get handed out for incremental gains, even if the core idea isn’t yours.
\nBut I think the secret is that the real leaderboard is the skills you build. The folks grinding on novel features, debugging leaks, or dissecting winners’ code. They’re the ones who’ll dominate the next comp. Clone wars might win short-term karma, but they don’t teach you anything.
\nStill sucks though. I wish originality carried more weight in medals.
", "votes": 3, "replies": [] }, { "id": 3249384, "author_name": "Gerwyn", "author_url": "", "post_date": "2025-07-16T11:13:22.217000", "content": "I can only speak for competition medals. IMO, it's harder to achieve than you'd imagine. Usually, you'll need to at the very least beat the public baseline, and with something different, or novel, enough to differentiate.
\nLB standing now is not really meaningful (well, excluding the gold zone obviously), since we have about 2 months to go. My observation is that the last 2 weeks are more crucial, as it's the window when GMs join, and when teams merge.
\nIn any case, i'll offer you a positive bit for high scoring public notebooks:
\nHigher scoring public notebooks -> more NPCs enter to fork and submit -> more spots for medals (since the number of spots is proportional to teams which made at least one submission)
1854 teams at time of writing, 0.84 notebook released 8 hours ago. you can probably download the LB data now, and check in 1-2 days to see the number of 0.84 LB with less than 10 subs maybe to see how much the notebook attracted new teams.
", "votes": 3, "replies": [ { "id": 3249388, "author_name": "Huseyin Gur", "author_url": "", "post_date": "2025-07-16T11:25:12.247000", "content": "Thanks for the detailed and helpful explanation — I appreciate it.
\nOne thing I’m still trying to understand is the actual value or impact of notebook medals.
\nBecause from what I’ve seen, some people make very minor changes to existing notebooks and still get gold medals and a lot of upvotes — even if they’ve done no real work themselves.
So I’m wondering, how meaningful are these medals?
", "votes": -1, "replies": [ { "id": 3249397, "author_name": "Gerwyn", "author_url": "", "post_date": "2025-07-16T12:00:04.137000", "content": "same viewpoint as @shlomoron - thats how kaggle work for competition notebooks at least: i am quite sure majority are just doing a sort by Public Score, so higher score notebook usually get more coverage in that sense.
Also, the recent update (i believe), gives upvote for every fork, so it's no surprise that high scores notebook have high correlation with upvote (more coverage + more forks).
", "votes": 1, "replies": [] } ] } ] }, { "id": 3249409, "author_name": "EliKal", "author_url": "", "post_date": "2025-07-16T12:20:14.833000", "content": "Public notebooks are just a helpful resource. Sure, people love seeing themselves in the top 200-300 on Kaggle, but real progress comes from getting inspired and grinding it out. I’d say my result is 50% vibe coding, 45% public notebooks, and 5% brainpower… well, maybe even less 😅
", "votes": 1, "replies": [ { "id": 3249612, "author_name": "Noah Stefancik", "author_url": "", "post_date": "2025-07-16T18:57:18.030000", "content": "Haha, I’ve been hearing a lot about vibe coding recently. What have you found works best for data science applications in terms of models and workflows? Just curious what high level people have found to work best! -Thanks
", "votes": 0, "replies": [ { "id": 3249630, "author_name": "EliKal", "author_url": "", "post_date": "2025-07-16T19:44:03.770000", "content": "I use different LLMs, designing with one, critiquing with another, debugging, and training. If one doesn’t work, I switch to another LLM. I draw a lot of inspiration from public notebooks. There are some best practices online 🙂
", "votes": 1, "replies": [] } ] } ] }, { "id": 3253279, "author_name": "Taimour Nazar", "author_url": "", "post_date": "2025-07-24T11:16:21.083000", "content": "1- Trust you CV score.
\n2- LB is a good measure of how you are doing but its not final.
\n3- View leaderboard of last CMI competition \"Problematic Internet Use\" you will see that the people on top had jumped more than 1000 or 2000 points after final scoring.
\n4- Don't worry about what people are doing. Just give your best.
Good luck
", "votes": 2, "replies": [] }, { "id": 3250938, "author_name": "@kaggleqrdl", "author_url": "", "post_date": "2025-07-19T14:12:33.357000", "content": "\nYes, they've basically made notebook medals without any value for the moment.
\nHow this will play out longer term:
\nUntil one of these things happens, notebook GM status is nonsense and is rapidly getting diluted. They can retroactively rug users again, but they don't like removing GM status and having a GM status without actually having it feels humiliating.
\nI've opted out of the progression tiers as they are pretty silly until it stabilizes or is fixed.
\nHonestly, I like the tweaked notebooks, read them all, and would publish some myself.
\nBut obv they are not gold medal worthy. And much higher value notebooks are being ignored because of it. If I didn't know better, I'd assume they are trying to destroy the noteboook progression tier.
", "votes": 2, "replies": [ { "id": 3250944, "author_name": "Huseyin Gur", "author_url": "", "post_date": "2025-07-19T14:36:47.773000", "content": "Absolutely agree, this captures the issue perfectly.
", "votes": 0, "replies": [] } ] }, { "id": 3249958, "author_name": "Fabio", "author_url": "", "post_date": "2025-07-17T12:44:20.207000", "content": "It's fine to accept that, in a past competition it was happening the same and look at the final results CMI problematic internet use.
\nAll the copied kernel will fall in the same pit at evaluation, effectively eliminating tons of submissions.
", "votes": 2, "replies": [] }, { "id": 3252790, "author_name": "Erigitola Wahab Olatunji", "author_url": "", "post_date": "2025-07-23T12:15:24.423000", "content": "This is a really important question—and one that many new (and even experienced) Kagglers ask at some point.
\nKaggle encourages sharing and collaboration, which is part of what makes it such a vibrant learning community. But yes, there's a fine line between building upon others’ work and cloning for easy gains. When a minor change to a high-performing notebook gets heavily upvoted or even rewarded with medals, it can feel unfair—especially to those investing serious time into original modeling or exploration.
", "votes": 1, "replies": [] }, { "id": 3249336, "author_name": "greySnow", "author_url": "", "post_date": "2025-07-16T09:02:54.860000", "content": "That's how Kaggle works.
\nYou may think it's stupid, but then try to think on the alternatives and whether they will incentivise Kagglers to share more or less of the helpful content.
I’m not against sharing. But getting a medal for changing one parameter in someone else’s notebook isn’t sharing.
\nI think this doesn’t encourage contribution — it encourages copy-paste.
", "votes": 1, "replies": [ { "id": 3249348, "author_name": "greySnow", "author_url": "", "post_date": "2025-07-16T09:35:08.457000", "content": "So, I ask again, what is your alternative?
\nOnly, before you answer, please remember that Kagglers are very good at gaming any system.
I get that any system can be gamed, but it still feels worth questioning when very small changes lead to big rewards.
\nAlso, I’m just a user sharing an observation — coming up with alternatives isn’t really my role.
", "votes": -2, "replies": [] }, { "id": 3249366, "author_name": "greySnow", "author_url": "", "post_date": "2025-07-16T10:17:20.890000", "content": "It's easy to criticize. But only criticism does not contribute much.
\nEspecially when it was heard so many times before.
\nYou have no idea how many times I have seen this observation already…I assure you that the Kaggle staff is aware.
\nUntil you or someone else comes up with a better alternative, this is the best we can have.
I understand, and I appreciate your perspective.
\nAs someone new here, I’m just trying to learn and make sense of how things work.
\nI think If this kind of feedback has been shared many times before, maybe it simply shows that some people feel strongly about it — which could be worth noting.
@shlomoron fork + submit of a public notebook without any changes or with exact same predictions should not be considered a valid submission and should not appear on the leaderboard.
\nThat is easy to implement for Kaggle, this does not penalize people who want to fork and change a few parameters, this still encourages people to share if they want to. But you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition). How about that proposal ?
", "votes": 1, "replies": [] }, { "id": 3249400, "author_name": "stefanoclss", "author_url": "", "post_date": "2025-07-16T12:07:11.700000", "content": "I don't see the issue, winning a Kaggle comp is not about changing a few params of a public notebook. It is the ability to merge it with your current pipeline and gain an advantage.
", "votes": 2, "replies": [] }, { "id": 3249425, "author_name": "", "author_url": "", "post_date": "2025-07-16T13:03:03.503000", "content": "", "votes": 0, "replies": [] }, { "id": 3249428, "author_name": "Huseyin Gur", "author_url": "", "post_date": "2025-07-16T13:03:39.743000", "content": "I really like the idea of excluding submissions with identical prediction files from the leaderboard — it’s a fair and effective solution.
\nThat said, a separate issue is users getting gold medals for copied notebooks with minimal effort (e.g., just changing a weight, N_SPLIT, or random_state).
Your proposal is quite useless, I will just add some extra lines of junk code so the code is not the same, and a very small noise to prediction (or will change the seed?).
\nIt adds complexity while solving nothing.
\n(And anyway it's not the issue we are talking about here- and if you ask me, it's not an issue at all).
\n\nBut you don't allow to mess up the leader board with literal copies (which is always the case for 200 people in each competition).
\n
There used to be a symbol \"< >\" on the LB to indicate which public notebook scored that on the LB (easier to ignore those repeated subs), though i havent seen it for a while so it may have been removed.
", "votes": 2, "replies": [] }, { "id": 3249621, "author_name": "Sheikh Muhammad Abdullah", "author_url": "", "post_date": "2025-07-16T19:22:46.580000", "content": "Did you see any gold medal solution write-up where someone says, \"I just changed the seed or splits and win a gold medal or silver\"?
\n@huseyingur, I haven't come across any like that.
\nyou have to work hard for gold, silver and winning.
There was. Place 5 in AIMO (the first one). He passed up the price after saying that he feel he does not deserve it.
\nBut it can happen only in conpetitions with high lottery factor.
This was a recent change. https://www.kaggle.com/discussions/product-announcements/588704
\nThey are probably hoping it will fix as the community self-policies and we are in a temporary gaming phase.
\nI'm 90% sure something will have to change as I mentioned above. Maybe 10 forks = 1 upvote or something. There are options.
\nI'm kinda hoping they will abandon the whole progression tier thing.
\nThe truth is the titles don't mean much to industry anymore as more specific skillsets, especially those which are LLM specific, are what is mostly sought after these days. Some comp wins are specific for certain verticals, but again, it's not the title. It's the comp you won.
\nI think removing the discussion tier really broke everything. They should have tried fixing it instead.
\nMaybe they should have got rid of the comp and notebook tier, perhaps and just kept discussions? Considering in discussions is where our colored circles actually show up, that makes more sense to me.
\nPeople who rank #1 in comps and don't participate much in the community, I be skeptical if they actually care about the titles.
", "votes": -1, "replies": [] }, { "id": 3251035, "author_name": "@kaggleqrdl", "author_url": "", "post_date": "2025-07-19T18:49:20.087000", "content": "Gold is non trivial, silver can be got by luck fairly easily. Consistently winning silver without piling up a lot of bronze can be harder.
\n\n\nDid you see any gold medal solution write-up where someone says, \"I just changed the seed or splits and win a gold medal or silver\"?
\n
\n @huseyingur, I haven't come across any like that.
\n you have to work hard for gold, silver and winning.
My first comp, I made a thoughtful, but not overly effortful, change to a notebook, and lucked out with a silver. In comps with poor cv / lb correlation that result in significant shakeups, this happens quite a bit. Luck plays a very big role in these comps, even for gold.
\nBut yes, consistently winning well is much harder for sure. You generally want to avoid bronze and even silver if you can.
", "votes": -1, "replies": [] } ] } ] }, { "id": 3249616, "author_name": "Sheikh Muhammad Abdullah", "author_url": "", "post_date": "2025-07-16T19:12:35.670000", "content": "I think sometimes luck might be on your side but just changing splits or seeds isn’t enough to win a silver or gold.
\nIn this competition specifically, most of the top public solutions are using saved models, so there's definitely more going on than just tuning seeds.
Well public notebooks are just a helpful resource if there is any doubt. Ofcourse, people do get distracted and opt an easier option by copying but real progress comes from getting inspired and grinding it out.
", "votes": 0, "replies": [] }, { "id": 3250849, "author_name": "Sarah Arshad", "author_url": "", "post_date": "2025-07-19T09:10:39.517000", "content": "Public notebooks are just a helpful resource.
", "votes": 0, "replies": [] }, { "id": 3249368, "author_name": "", "author_url": "", "post_date": "2025-07-16T10:19:43.323000", "content": "", "votes": -3, "replies": [ { "id": 3249382, "author_name": "", "author_url": "", "post_date": "2025-07-16T11:09:16.387000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3249312": "Someone takes an existing top notebook, changes N_SPLITS from 5 to 10 or weights, reruns it, gets +0.01 LB boost, and publishes it.\n\nThat clone gets 100+ upvotes or even a gold medal. People who’ve done no actual work in the competition just run these notebooks and end up high on the leaderboard.\n\nWhat’s the point of this system? I’m new here, just trying to understand.", "3249962": "Check out the private leaderboard from a previous competition by the same host to see how much it can differ from the public leaderboard: https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard\nIn that competition, the 1st place solution increased by 1,488 ranks, and the 2nd place increased by 2,603. Many Kaggle competitions end with a significant public LB shake-up like this.\n\nThen there are some competitions where free loaders end up with medals. A recent example is the March Machine Learning Mania 2025 competition: https://www.kaggle.com/competitions/march-machine-learning-mania-2025 \nIn that one, 69 notebooks had the exact same LB score, most likely from copying the same 2 public notebooks. But here's the thing: over 3,000 data scientists followed that competition, and many probably noticed. So even if someone got a silver, it came with doubt and reputation risk...\n\nBottom line: take the public leaderboard with a grain of salt. Avoid using public approaches you don’t fully understand or trust. And most importantly, rely on your own testing scheme rather than chasing small public LB gains. IMO most importantly kaggle is a platform for practice and learning and medals are nice little encouragements. If you aren't working it out yourself, you aren't improving your expertise, then what is the point of participating anyway. ", "3249398": "But this way you cannot win gold, extremely merge and tune public notebook just can let you win silver and no one will care the final solution you write. And most important thing, you do not learn anything in this competition. How you discover the relationship between these physical measurements and the gestures is what you need to learn and gain experience.\n", "3251098": "Totally get the frustration; here’s what’s really happening:\n\n**Why tiny tweaks + reposts get traction**\n\n* Early visibility + recency bias: a freshly published notebook with a small LB bump surfaces in “Recently Published” and people auto‑upvote.\n* Many newcomers want a ready notebook to fork; they reward anything that *runs* and shows +0.01, even if it’s just noise.\n* A 0.01 public LB gain often *is* variance (different seed / KFold split) rather than real improvement.\n\n**Does this hurt the competition?**\nNot really. Public LB ≠ final ranking. Shallow clones rarely hold once the private LB (or second stage) scores drop. Serious competitors rely on local CV, not public noise.\n\n**How to spot real contribution**\n\n* Clear, reasoned CV strategy (stratification rationale, leak checks, stability stats).\n* Novel features / preprocessing / architecture changes explained and *validated* (tables of fold scores, seed averages).\n* Error analysis, robustness checks, inference speedups, ablation studies.\n* Reproducibility: fixed seeds, version pinning, concise modular code.\n\n**Why clones still appear**\n\n* Low barrier to entry encourages learning by modification (not evil, just low signal).\n* Medals reflect aggregate community voting; the system can’t perfectly judge originality.\n\n**What to do instead of worrying**\n\n* Ignore superficial forks; focus on building a rock‑solid offline CV that predicts *both* public + private.\n* Track your seeds; average scores; log fold dispersion.\n* Run ablations: prove each claimed improvement.\n* Share *one* well‑documented notebook late enough (or with hidden tricks locally) if you care about protecting ideas.\n* Use Discussions for deeper write‑ups (methodology > code dump). Those attract informed voters.\n\n**Improving your own signal**\n\n* Publish: (1) Clean baseline with robust CV; (2) Feature / model expansion with ablations; (3) Final inference / ensemble write‑up.\n* Include: dataset issues found, leakage tests, outlier handling, model diagnostics.\n* Time your publication when noise is lower (not right after launch surge).\n\n**Sanity checks before praising a “+0.01” notebook**\n\n* Compare its CV vs its claimed LB jump.\n* Look for seed averaging; absence often means variance exploitation.\n* Diff the code: are changes substantive or just N\\_SPLITS, learning rate, seed?\n\n**Mindset**\nTreat public LB improvements under \\~0.02 as “unconfirmed” until replicated in your CV. Your edge comes from rigor, not chasing every micro‑gain someone posts.\n\nIn short: the system surfaces noise, but it doesn’t decide the final outcome. Build real skill: solid validation + genuine innovation. Let the clones farm early upvotes; they almost never win medals that matter (final private LB).\n\nWelcome aboard—focus on craft, you’ll rise past the noise.", "3252004": "I totally feel you! it’s frustrating to see low-effort clones rocket up the LB while original work gets buried. \n\nFrom my experience, there is a messy reality that covers the good , the bad and the ugly. \n\nThe Good! Tiny tweaks can legitimately improve models (say more folds = better CV = smarter stacking). Sometimes that +0.01 is the difference between 1st and 10th.\n\nThen there is the Bad. Most of these are just noise. People cargo-culting changes without understanding why they work (or don’t). Upvotes often mean “this gave me free LB points,” not “this is insightful.”\n\nAnd then finally the Ugly. Kaggle’s reward system does incentivize this. Medals get handed out for incremental gains, even if the core idea isn’t yours.\n\nBut I think the secret is that the real leaderboard is the skills you build. The folks grinding on novel features, debugging leaks, or dissecting winners’ code. They’re the ones who’ll dominate the next comp. Clone wars might win short-term karma, but they don’t teach you anything.\n\nStill sucks though. I wish originality carried more weight in medals.", "3249384": "I can only speak for competition medals. IMO, it's harder to achieve than you'd imagine. Usually, you'll need to at the very least beat the public baseline, and with something different, or novel, enough to differentiate. \n\nLB standing now is not really meaningful (well, excluding the gold zone obviously), since we have about 2 months to go. My observation is that the last 2 weeks are more crucial, as it's the window when GMs join, and when teams merge.\n\nIn any case, i'll offer you a positive bit for high scoring public notebooks: \nHigher scoring public notebooks -> more NPCs enter to fork and submit -> more spots for medals (since the number of spots is proportional to teams which made at least one submission)\n\n1854 teams at time of writing, 0.84 notebook released 8 hours ago. you can probably download the LB data now, and check in 1-2 days to see the number of 0.84 LB with less than 10 subs maybe to see how much the notebook attracted new teams.", "3249409": "Public notebooks are just a helpful resource. Sure, people love seeing themselves in the top 200-300 on Kaggle, but real progress comes from getting inspired and grinding it out. I’d say my result is 50% vibe coding, 45% public notebooks, and 5% brainpower... well, maybe even less 😅", "3253279": "1- Trust you CV score.\n2- LB is a good measure of how you are doing but its not final.\n3- View leaderboard of last CMI competition \"Problematic Internet Use\" you will see that the people on top had jumped more than 1000 or 2000 points after final scoring.\n4- Don't worry about what people are doing. Just give your best.\n\nGood luck", "3250938": "@mylesoneill \n\nYes, they've basically made notebook medals without any value for the moment. \n\nHow this will play out longer term:\n\n1. People will stop forking these tweaked forks (unlikely, and sort of irritating to have to avoid forking a notebook)\n2. Kaggle will warn people for making forks without huge improvements (it's actually against the guidelines, believe it or not, though I really don't know why. It's the basis of opensource and these small improvements when added up in totality are often quite useful!! The whole concept of opensource is leveraging the crowd for small contributions)\n3. Kaggle will increase # of notebook medals (really bad idea - as it punishes people who make thoughtful notebooks)\n4. Kaggle will increase the level for fork votes to master or gm level (my fav option)\n5. Kaggle will remove the notebook progession tier\n6. Kaggle will tweak it in some other way\n\nUntil one of these things happens, notebook GM status is nonsense and is rapidly getting diluted. They can retroactively rug users again, but they don't like removing GM status and having a GM status without actually having it feels humiliating.\n\nI've opted out of the progression tiers as they are pretty silly until it stabilizes or is fixed.\n\nHonestly, I like the tweaked notebooks, read them all, and would publish some myself. \n\nBut obv they are not gold medal worthy. And much higher value notebooks are being ignored because of it. If I didn't know better, I'd assume they are trying to destroy the noteboook progression tier.", "3249958": "It's fine to accept that, in a past competition it was happening the same and look at the final results [CMI problematic internet use](https://www.kaggle.com/competitions/child-mind-institute-problematic-internet-use/leaderboard).\n\nAll the copied kernel will fall in the same pit at evaluation, effectively eliminating tons of submissions. \n\n", "3252790": "This is a really important question—and one that many new (and even experienced) Kagglers ask at some point.\n\nKaggle encourages sharing and collaboration, which is part of what makes it such a vibrant learning community. But yes, there's a fine line between building upon others’ work and cloning for easy gains. When a minor change to a high-performing notebook gets heavily upvoted or even rewarded with medals, it can feel unfair—especially to those investing serious time into original modeling or exploration.", "3249336": "That's how Kaggle works. \nYou may think it's stupid, but then try to think on the alternatives and whether they will incentivise Kagglers to share more or less of the **helpful** content. ", "3249616": "I think sometimes luck might be on your side but just changing splits or seeds isn’t enough to win a silver or gold.\nIn this competition specifically, most of the top public solutions are using saved models, so there's definitely more going on than just tuning seeds. ", "3251012": "Well public notebooks are just a helpful resource if there is any doubt. Ofcourse, people do get distracted and opt an easier option by copying but real progress comes from getting inspired and grinding it out.", "3250849": "Public notebooks are just a helpful resource.", "3249368": "" } }