Recently, an upsetting kernel has been published : \nhttps://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string (EDIT)
\n\nIt basically allows us to reuse previous submissions into a string. Which can lead to a significant boost in performance.\nAs this is not recognized as external data but still kinda is, can anyone (from kaggle?) confirm that such practices are forbidden (or not but I would be surprised) ?
\n\nThanks in advance
", "messages": [ { "id": 457753, "postDate": "2019-01-18T03:14:08.927Z", "content": "The final score will be tested on a new private test set, this should not be a big problem. BTW, this is cheating, indeed
", "rawMarkdown": "The final score will be tested on a new private test set, this should not be a big problem. BTW, this is cheating, indeed", "votes": 10, "replies": [ { "id": 457868, "postDate": "2019-01-18T08:34:03.147Z", "content": "Still, having more labelled samples can be useful for pseudo-labelling and can improve a model.
", "rawMarkdown": "Still, having more labelled samples can be useful for pseudo-labelling and can improve a model." } ] }, { "id": 457527, "postDate": "2019-01-17T16:15:53.090Z", "content": "Recently, an upsetting kernel has been published : \nhttps://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string (EDIT)
\n\nIt basically allows us to reuse previous submissions into a string. Which can lead to a significant boost in performance.\nAs this is not recognized as external data but still kinda is, can anyone (from kaggle?) confirm that such practices are forbidden (or not but I would be surprised) ?
\n\nThanks in advance
", "rawMarkdown": "Recently, an upsetting kernel has been published : \nhttps://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string (EDIT)\n\nIt basically allows us to reuse previous submissions into a string. Which can lead to a significant boost in performance.\nAs this is not recognized as external data but still kinda is, can anyone (from kaggle?) confirm that such practices are forbidden (or not but I would be surprised) ?\n\nThanks in advance", "votes": 9 }, { "id": 457878, "postDate": "2019-01-18T08:47:22.967Z", "content": "This is an insincere kernel :)
", "rawMarkdown": "This is an insincere kernel :)", "votes": 5 }, { "id": 457904, "postDate": "2019-01-18T10:08:50.963Z", "content": "@theoviel, don't worry about it as this is clearly against the rule.
\n\nThe whole idea of a kernel only competition is that you run your model end-to-end within the kernel environment with time constraint and all. I just finished competing in the \"Travelling Santa 2018 - prime paths\" contest and a lot of similar questions were asked about the kernel prize part. The Kaggle team made it very clear that you have to do all your pre-processing plus modelling within the kernel environment.
", "rawMarkdown": "@theoviel, don't worry about it as this is clearly against the rule.\n\nThe whole idea of a kernel only competition is that you run your model end-to-end within the kernel environment with time constraint and all. I just finished competing in the [\"Travelling Santa 2018 - prime paths\"][1] contest and a lot of similar questions were asked about the kernel prize part. The Kaggle team made it very clear that you have to do all your pre-processing plus modelling within the kernel environment.\n\n\n [1]: https://www.kaggle.com/c/traveling-santa-2018-prime-paths", "votes": 2, "replies": [ { "id": 458300, "postDate": "2019-01-19T09:43:09.143Z", "content": "I think there are several problems if someone wants to cheat.
\n\nSuppose that I did fixing misspelling words out side of Kaggle kernel environment. Let say I have a dictionary including 10,000 fixed words. If I directly copy my dictionary into kernel, I bet the kernel will be crashed. So, using this trick I think I can have an advantage for cleaning data. Of course, this data does not depend on testset. It should be no problem on phase 2.
\nIt is a simple case. Let think about if someone wants to bring their \"pretrained weight\" into their models ;).
I wonder if Kaggle team can review all the kernel codes to know if they are cheating or not
", "rawMarkdown": "I think there are several problems if someone wants to cheat. \n\nSuppose that I did fixing misspelling words out side of Kaggle kernel environment. Let say I have a dictionary including 10,000 fixed words. If I directly copy my dictionary into kernel, I bet the kernel will be crashed. So, using this trick I think I can have an advantage for cleaning data. Of course, this data does not depend on testset. It should be no problem on phase 2. \nIt is a simple case. Let think about if someone wants to bring their \"pretrained weight\" into their models ;). \n\nI wonder if Kaggle team can review all the kernel codes to know if they are cheating or not", "votes": 1 }, { "id": 458361, "postDate": "2019-01-19T12:49:36.810Z", "content": "Your pre-trained weights will be considered an external data hence against the rule.
\n\nOn your other point, Kaggle go through cheaters removal at the end of every competition before finalizing the results. I remember in my 1st Kaggle competition, I jumped 2 percentage points i.e. from top 15% to 13% a few days after the closing. So food for thought.
", "rawMarkdown": "Your pre-trained weights will be considered an external data hence against the rule.\n\nOn your other point, Kaggle go through cheaters removal at the end of every competition before finalizing the results. I remember in my 1st Kaggle competition, I jumped 2 percentage points i.e. from top 15% to 13% a few days after the closing. So food for thought." } ] }, { "id": 457537, "postDate": "2019-01-17T16:48:58.243Z", "content": "I think you mean this kernel: https://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string
", "rawMarkdown": "I think you mean this kernel: https://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string", "votes": 2, "replies": [ { "id": 457662, "postDate": "2019-01-17T22:05:49.593Z", "content": "Oops yeah, messed up my c/c !
", "rawMarkdown": "Oops yeah, messed up my c/c !" } ] }, { "id": 457563, "postDate": "2019-01-17T18:09:50.083Z", "content": "Well it is pretty obvious that this is cheating so I would not worry too much about it.
", "rawMarkdown": "Well it is pretty obvious that this is cheating so I would not worry too much about it.", "votes": -1 }, { "id": 457898, "postDate": "2019-01-18T09:55:27.677Z", "content": "Training your model locally and submitting only indexes saves a lot of time, so it is not cheating. Using external information in the final kernel, such as pseudo labeled public test data is cheating.
", "rawMarkdown": "Training your model locally and submitting only indexes saves a lot of time, so it is not cheating. Using external information in the final kernel, such as pseudo labeled public test data is cheating.", "votes": -5 }, { "id": 457538, "postDate": "2019-01-17T16:49:39.847Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 457753, "author_name": "good good study", "author_url": "", "post_date": "2019-01-18T03:14:08.927000", "content": "The final score will be tested on a new private test set, this should not be a big problem. BTW, this is cheating, indeed
", "votes": 10, "replies": [ { "id": 457868, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-01-18T08:34:03.147000", "content": "Still, having more labelled samples can be useful for pseudo-labelling and can improve a model.
", "votes": 0, "replies": [] } ] }, { "id": 457878, "author_name": "Juju", "author_url": "", "post_date": "2019-01-18T08:47:22.967000", "content": "This is an insincere kernel :)
", "votes": 5, "replies": [] }, { "id": 457904, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2019-01-18T10:08:50.963000", "content": "@theoviel, don't worry about it as this is clearly against the rule.
\n\nThe whole idea of a kernel only competition is that you run your model end-to-end within the kernel environment with time constraint and all. I just finished competing in the \"Travelling Santa 2018 - prime paths\" contest and a lot of similar questions were asked about the kernel prize part. The Kaggle team made it very clear that you have to do all your pre-processing plus modelling within the kernel environment.
", "votes": 2, "replies": [ { "id": 458300, "author_name": "cab", "author_url": "", "post_date": "2019-01-19T09:43:09.143000", "content": "I think there are several problems if someone wants to cheat.
\n\nSuppose that I did fixing misspelling words out side of Kaggle kernel environment. Let say I have a dictionary including 10,000 fixed words. If I directly copy my dictionary into kernel, I bet the kernel will be crashed. So, using this trick I think I can have an advantage for cleaning data. Of course, this data does not depend on testset. It should be no problem on phase 2.
\nIt is a simple case. Let think about if someone wants to bring their \"pretrained weight\" into their models ;).
I wonder if Kaggle team can review all the kernel codes to know if they are cheating or not
", "votes": 1, "replies": [] }, { "id": 458361, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2019-01-19T12:49:36.810000", "content": "Your pre-trained weights will be considered an external data hence against the rule.
\n\nOn your other point, Kaggle go through cheaters removal at the end of every competition before finalizing the results. I remember in my 1st Kaggle competition, I jumped 2 percentage points i.e. from top 15% to 13% a few days after the closing. So food for thought.
", "votes": 0, "replies": [] } ] }, { "id": 457537, "author_name": "Mark Worrall", "author_url": "", "post_date": "2019-01-17T16:48:58.243000", "content": "I think you mean this kernel: https://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string
", "votes": 2, "replies": [ { "id": 457662, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-01-17T22:05:49.593000", "content": "Oops yeah, messed up my c/c !
", "votes": 0, "replies": [] } ] }, { "id": 457563, "author_name": "Psi", "author_url": "", "post_date": "2019-01-17T18:09:50.083000", "content": "Well it is pretty obvious that this is cheating so I would not worry too much about it.
", "votes": -1, "replies": [] }, { "id": 457898, "author_name": "", "author_url": "", "post_date": "2019-01-18T09:55:27.677000", "content": "Training your model locally and submitting only indexes saves a lot of time, so it is not cheating. Using external information in the final kernel, such as pseudo labeled public test data is cheating.
", "votes": -5, "replies": [] }, { "id": 457538, "author_name": "", "author_url": "", "post_date": "2019-01-17T16:49:39.847000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "457753": "The final score will be tested on a new private test set, this should not be a big problem. BTW, this is cheating, indeed", "457527": "Recently, an upsetting kernel has been published : \nhttps://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string (EDIT)\n\nIt basically allows us to reuse previous submissions into a string. Which can lead to a significant boost in performance.\nAs this is not recognized as external data but still kinda is, can anyone (from kaggle?) confirm that such practices are forbidden (or not but I would be surprised) ?\n\nThanks in advance", "457878": "This is an insincere kernel :)", "457904": "@theoviel, don't worry about it as this is clearly against the rule.\n\nThe whole idea of a kernel only competition is that you run your model end-to-end within the kernel environment with time constraint and all. I just finished competing in the [\"Travelling Santa 2018 - prime paths\"][1] contest and a lot of similar questions were asked about the kernel prize part. The Kaggle team made it very clear that you have to do all your pre-processing plus modelling within the kernel environment.\n\n\n [1]: https://www.kaggle.com/c/traveling-santa-2018-prime-paths", "457537": "I think you mean this kernel: https://www.kaggle.com/paulorzp/compressing-a-binary-submission-within-a-string", "457563": "Well it is pretty obvious that this is cheating so I would not worry too much about it.", "457898": "Training your model locally and submitting only indexes saves a lot of time, so it is not cheating. Using external information in the final kernel, such as pseudo labeled public test data is cheating.", "457538": "" } }