{ "id": 79547, "title": "0.786 HOW IS THIS POSSIBLE?!?!!?", "url": "/competitions/quora-insincere-questions-classification/discussion/79547", "author_name": "Mekaveli.", "post_date": "2019-02-05T11:09:16.141000", "votes": 17, "comment_count": 35, "views": 0, "content": "

Looks like people from different worlds zzz. I'm so looking forward to learn from them after competition ends.

", "messages": [ { "id": 466538, "postDate": "2019-02-05T15:26:27.457Z", "content": "

Let's wait until the second stage results :)

", "rawMarkdown": "Let's wait until the second stage results :)", "votes": 24, "replies": [ { "id": 466673, "postDate": "2019-02-05T19:50:20.663Z", "content": "

Ну-с, подождем-с. Опубликуйте, потом решение, пожалуйста (с Бертами там или без :))

", "rawMarkdown": "Ну-с, подождем-с. Опубликуйте, потом решение, пожалуйста (с Бертами там или без :))", "votes": -10 }, { "id": 466878, "postDate": "2019-02-06T06:09:55.260Z", "content": "

pls use Engish\nfor those worried: no BERT was used

", "rawMarkdown": "pls use Engish\nfor those worried: no BERT was used ", "votes": 19 }, { "id": 466889, "postDate": "2019-02-06T06:29:39.143Z", "content": "

As expected from you!

", "rawMarkdown": "As expected from you!" }, { "id": 467042, "postDate": "2019-02-06T11:25:37.577Z", "content": "

Respect.

", "rawMarkdown": "Respect." }, { "id": 467132, "postDate": "2019-02-06T13:57:24.597Z", "content": "

fast ai with ULMfit? I don't know if you can get it without using internet and i got the idea on the last day of the competition. I saw that you used fast ai before on your github that's how i got the idea

", "rawMarkdown": "fast ai with ULMfit? I don't know if you can get it without using internet and i got the idea on the last day of the competition. I saw that you used fast ai before on your github that's how i got the idea" }, { "id": 471261, "postDate": "2019-02-14T08:09:20.733Z", "content": "

Despite overfit, I would still like to know how you achieved this score. :)

", "rawMarkdown": "Despite overfit, I would still like to know how you achieved this score. :)", "votes": 1 }, { "id": 471463, "postDate": "2019-02-14T14:01:38.603Z", "content": "

Pavel, could you please explain still what was it? Sorry about your final result

", "rawMarkdown": "Pavel, could you please explain still what was it? Sorry about your final result", "votes": 1 } ] }, { "id": 466428, "postDate": "2019-02-05T11:09:16.143Z", "content": "

Looks like people from different worlds zzz. I'm so looking forward to learn from them after competition ends.

", "rawMarkdown": "Looks like people from different worlds zzz. I'm so looking forward to learn from them after competition ends.", "votes": 15 }, { "id": 466462, "postDate": "2019-02-05T12:17:54.647Z", "content": "

I think guys just play game - they use some external data for score on this stage (to practice and scare us :)), but for second stage they will use different model.

", "rawMarkdown": "I think guys just play game - they use some external data for score on this stage (to practice and scare us :)), but for second stage they will use different model.", "votes": 13 }, { "id": 471028, "postDate": "2019-02-13T23:33:33.567Z", "content": "

Narrator: It wasn't possible.

", "rawMarkdown": "Narrator: It wasn't possible.", "votes": 8 }, { "id": 466433, "postDate": "2019-02-05T11:21:54.517Z", "content": "

Everybody's curious. However, they do have a right not to tell anybody. Anyway, you won't know until medals are distributed :)

", "rawMarkdown": "Everybody's curious. However, they do have a right not to tell anybody. Anyway, you won't know until medals are distributed :)", "votes": 6 }, { "id": 471258, "postDate": "2019-02-14T08:01:03.730Z", "content": "

You could get all the right answers for stage 1 by using Google search API. Toxic questions are not posted, non-toxic questions are on Quora.

", "rawMarkdown": "You could get all the right answers for stage 1 by using Google search API. Toxic questions are not posted, non-toxic questions are on Quora.", "votes": 3 }, { "id": 466863, "postDate": "2019-02-06T05:18:44.393Z", "content": "

I could have gotten 0.99 but I held back. I could have labeled 56k questions in 3 days :)

", "rawMarkdown": "I could have gotten 0.99 but I held back. I could have labeled 56k questions in 3 days :)", "votes": 4, "replies": [ { "id": 466877, "postDate": "2019-02-06T06:08:47Z", "content": "

labels are very noisy, doesn't seem possible

", "rawMarkdown": "labels are very noisy, doesn't seem possible", "votes": 7 } ] }, { "id": 466443, "postDate": "2019-02-05T11:54:32.180Z", "content": "

Somebody had shared that using BERT, this score is possible.

", "rawMarkdown": "Somebody had shared that using BERT, this score is possible.", "votes": 4, "replies": [ { "id": 466459, "postDate": "2019-02-05T12:12:46.177Z", "content": "

I don't think that they would have used BERT...Really looking forward to the solution by them, if they share...

", "rawMarkdown": "I don't think that they would have used BERT...Really looking forward to the solution by them, if they share..." }, { "id": 466467, "postDate": "2019-02-05T12:34:42.513Z", "content": "

Yeah, and you can use BERT + ensembles on your local machine and fill the submission file manually in the Kaggle kernel so that you can submit it to the LB. It won't work in stage 2 but it does show you how well your solution works on the stage 1 LB.

", "rawMarkdown": "Yeah, and you can use BERT + ensembles on your local machine and fill the submission file manually in the Kaggle kernel so that you can submit it to the LB. It won't work in stage 2 but it does show you how well your solution works on the stage 1 LB.", "votes": 2 }, { "id": 466482, "postDate": "2019-02-05T13:19:23.493Z", "content": "

Using bert violate the rule, doesn't it? \nI guess they dont use bert

", "rawMarkdown": "Using bert violate the rule, doesn't it? \nI guess they dont use bert", "votes": 1 }, { "id": 466489, "postDate": "2019-02-05T13:35:43.590Z", "content": "

Yes, I believe it violates the rules, the same way it is not allowed to use other embeddings than provided ones ... it is not allowed to use BERT transfer learning as it is not one of input sources

", "rawMarkdown": "Yes, I believe it violates the rules, the same way it is not allowed to use other embeddings than provided ones ... it is not allowed to use BERT transfer learning as it is not one of input sources" } ] }, { "id": 467047, "postDate": "2019-02-06T11:36:51.677Z", "content": "

My guess is you somehow derived the original ML-model quora used for labeling.

", "rawMarkdown": "My guess is you somehow derived the original ML-model quora used for labeling. ", "votes": 3, "replies": [ { "id": 468215, "postDate": "2019-02-08T14:06:59.207Z", "content": "

I guess the data used in stage 2 was real data like the training data. hopes not, otherwise the public leaderboard was just a joke.

", "rawMarkdown": "I guess the data used in stage 2 was real data like the training data. hopes not, otherwise the public leaderboard was just a joke." }, { "id": 468373, "postDate": "2019-02-08T19:21:33.603Z", "content": "

What do you mean? Public test data also was \"real\" data.

", "rawMarkdown": "What do you mean? Public test data also was \"real\" data." }, { "id": 469061, "postDate": "2019-02-10T12:10:59.383Z", "content": "

@philippsinger\nI guess the data used by the public lb is generated by certain rules (to prevent data leakage), while the data used by the private lb is part of real-world data. It’s just some guesswork.

", "rawMarkdown": "@philippsinger\nI guess the data used by the public lb is generated by certain rules (to prevent data leakage), while the data used by the private lb is part of real-world data. It’s just some guesswork." }, { "id": 469063, "postDate": "2019-02-10T12:13:07.953Z", "content": "

@xiaobai1123q You could google PLB data, so I doubt that it is different.

", "rawMarkdown": "@xiaobai1123q You could google PLB data, so I doubt that it is different." }, { "id": 471062, "postDate": "2019-02-14T01:25:27.527Z", "content": "

@philippsinger\nbut we could not know the data used in Private LB.

", "rawMarkdown": "@philippsinger\nbut we could not know the data used in Private LB.", "votes": -1 } ] }, { "id": 471552, "postDate": "2019-02-14T15:45:20.783Z", "content": "

Well, this is not possible after all...

\n\n

They dropped to 1027 position. I hope it is not because of a silly bug in their code.

", "rawMarkdown": "Well, this is not possible after all...\n\nThey dropped to 1027 position. I hope it is not because of a silly bug in their code.", "votes": 1 }, { "id": 471272, "postDate": "2019-02-14T08:29:06.153Z", "content": "

Nice trolling, Pavel! :)

", "rawMarkdown": "Nice trolling, Pavel! :)", "votes": 1 }, { "id": 471201, "postDate": "2019-02-14T06:22:06.853Z", "content": "

@icemekaveli, the shake up answered your question. You jumped higher than that score from 335 position. Congrats.

", "rawMarkdown": "@icemekaveli, the shake up answered your question. You jumped higher than that score from 335 position. Congrats.", "votes": 1, "replies": [ { "id": 471862, "postDate": "2019-02-15T02:09:19.603Z", "content": "

Thanks. You score nice too. Congrats.

", "rawMarkdown": "Thanks. You score nice too. Congrats.", "votes": 1 } ] }, { "id": 466879, "postDate": "2019-02-06T06:11:13.463Z", "content": "

I don't think they manually labelled submission file, then result would be higher :) Probably we should wait until stage 2 end.

", "rawMarkdown": "I don't think they manually labelled submission file, then result would be higher :) Probably we should wait until stage 2 end.", "votes": 1 }, { "id": 468313, "postDate": "2019-02-08T16:55:36.953Z", "content": "

Wouwww, good prize :)

", "rawMarkdown": "Wouwww, good prize :)" }, { "id": 466508, "postDate": "2019-02-05T14:22:57.107Z", "content": "

I hope we can learn it after the competition.

", "rawMarkdown": "I hope we can learn it after the competition. " }, { "id": 466490, "postDate": "2019-02-05T13:36:47.533Z", "content": "

Yeah I came to 0.695 after enormous work. Must be a different architechure. Looking forward to learn from them, if they can share...

", "rawMarkdown": "Yeah I came to 0.695 after enormous work. Must be a different architechure. Looking forward to learn from them, if they can share..." }, { "id": 466566, "postDate": "2019-02-05T16:26:08.987Z", "rawMarkdown": "", "votes": -1, "isDeleted": true }, { "id": 466504, "postDate": "2019-02-05T14:01:29.770Z", "content": "

Truly impressing, they are stratospheric geese, congrats to the team!\nI guess they found something, maybe simple, that we all missed while we were having fun with Neural Nets.\nI am looking forward to learn from them, I hope they will publish their solution.

", "rawMarkdown": "Truly impressing, they are stratospheric geese, congrats to the team!\nI guess they found something, maybe simple, that we all missed while we were having fun with Neural Nets.\nI am looking forward to learn from them, I hope they will publish their solution.", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 466538, "author_name": "", "author_url": "", "post_date": "2019-02-05T15:26:27.457000", "content": "

Let's wait until the second stage results :)

", "votes": 24, "replies": [ { "id": 466673, "author_name": "Serge", "author_url": "", "post_date": "2019-02-05T19:50:20.663000", "content": "

Ну-с, подождем-с. Опубликуйте, потом решение, пожалуйста (с Бертами там или без :))

", "votes": -10, "replies": [] }, { "id": 466878, "author_name": "", "author_url": "", "post_date": "2019-02-06T06:09:55.260000", "content": "

pls use Engish\nfor those worried: no BERT was used

", "votes": 19, "replies": [] }, { "id": 466889, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2019-02-06T06:29:39.143000", "content": "

As expected from you!

", "votes": 0, "replies": [] }, { "id": 467042, "author_name": "guiem", "author_url": "", "post_date": "2019-02-06T11:25:37.577000", "content": "

Respect.

", "votes": 0, "replies": [] }, { "id": 467132, "author_name": "Anas Charafi", "author_url": "", "post_date": "2019-02-06T13:57:24.597000", "content": "

fast ai with ULMfit? I don't know if you can get it without using internet and i got the idea on the last day of the competition. I saw that you used fast ai before on your github that's how i got the idea

", "votes": 0, "replies": [] }, { "id": 471261, "author_name": "Rahul Agarwal", "author_url": "", "post_date": "2019-02-14T08:09:20.733000", "content": "

Despite overfit, I would still like to know how you achieved this score. :)

", "votes": 1, "replies": [] }, { "id": 471463, "author_name": "Anna Novikova", "author_url": "", "post_date": "2019-02-14T14:01:38.603000", "content": "

Pavel, could you please explain still what was it? Sorry about your final result

", "votes": 1, "replies": [] } ] }, { "id": 466462, "author_name": "Serge", "author_url": "", "post_date": "2019-02-05T12:17:54.647000", "content": "

I think guys just play game - they use some external data for score on this stage (to practice and scare us :)), but for second stage they will use different model.

", "votes": 13, "replies": [] }, { "id": 471028, "author_name": "Peter Hurford", "author_url": "", "post_date": "2019-02-13T23:33:33.567000", "content": "

Narrator: It wasn't possible.

", "votes": 8, "replies": [] }, { "id": 466433, "author_name": "Yury Kashnitsky", "author_url": "", "post_date": "2019-02-05T11:21:54.517000", "content": "

Everybody's curious. However, they do have a right not to tell anybody. Anyway, you won't know until medals are distributed :)

", "votes": 6, "replies": [] }, { "id": 471258, "author_name": "Artyom Palvelev", "author_url": "", "post_date": "2019-02-14T08:01:03.730000", "content": "

You could get all the right answers for stage 1 by using Google search API. Toxic questions are not posted, non-toxic questions are on Quora.

", "votes": 3, "replies": [] }, { "id": 466863, "author_name": "Kepler456b", "author_url": "", "post_date": "2019-02-06T05:18:44.393000", "content": "

I could have gotten 0.99 but I held back. I could have labeled 56k questions in 3 days :)

", "votes": 4, "replies": [ { "id": 466877, "author_name": "", "author_url": "", "post_date": "2019-02-06T06:08:47", "content": "

labels are very noisy, doesn't seem possible

", "votes": 7, "replies": [] } ] }, { "id": 466443, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2019-02-05T11:54:32.180000", "content": "

Somebody had shared that using BERT, this score is possible.

", "votes": 4, "replies": [ { "id": 466459, "author_name": "Rahul Agarwal", "author_url": "", "post_date": "2019-02-05T12:12:46.177000", "content": "

I don't think that they would have used BERT...Really looking forward to the solution by them, if they share...

", "votes": 0, "replies": [] }, { "id": 466467, "author_name": "Fredrik Jonsson", "author_url": "", "post_date": "2019-02-05T12:34:42.513000", "content": "

Yeah, and you can use BERT + ensembles on your local machine and fill the submission file manually in the Kaggle kernel so that you can submit it to the LB. It won't work in stage 2 but it does show you how well your solution works on the stage 1 LB.

", "votes": 2, "replies": [] }, { "id": 466482, "author_name": "hahaha", "author_url": "", "post_date": "2019-02-05T13:19:23.493000", "content": "

Using bert violate the rule, doesn't it? \nI guess they dont use bert

", "votes": 1, "replies": [] }, { "id": 466489, "author_name": "A.Barqawi", "author_url": "", "post_date": "2019-02-05T13:35:43.590000", "content": "

Yes, I believe it violates the rules, the same way it is not allowed to use other embeddings than provided ones ... it is not allowed to use BERT transfer learning as it is not one of input sources

", "votes": 0, "replies": [] } ] }, { "id": 467047, "author_name": "Dieter", "author_url": "", "post_date": "2019-02-06T11:36:51.677000", "content": "

My guess is you somehow derived the original ML-model quora used for labeling.

", "votes": 3, "replies": [ { "id": 468215, "author_name": "Bai", "author_url": "", "post_date": "2019-02-08T14:06:59.207000", "content": "

I guess the data used in stage 2 was real data like the training data. hopes not, otherwise the public leaderboard was just a joke.

", "votes": 0, "replies": [] }, { "id": 468373, "author_name": "Psi", "author_url": "", "post_date": "2019-02-08T19:21:33.603000", "content": "

What do you mean? Public test data also was \"real\" data.

", "votes": 0, "replies": [] }, { "id": 469061, "author_name": "Bai", "author_url": "", "post_date": "2019-02-10T12:10:59.383000", "content": "

@philippsinger\nI guess the data used by the public lb is generated by certain rules (to prevent data leakage), while the data used by the private lb is part of real-world data. It’s just some guesswork.

", "votes": 0, "replies": [] }, { "id": 469063, "author_name": "Psi", "author_url": "", "post_date": "2019-02-10T12:13:07.953000", "content": "

@xiaobai1123q You could google PLB data, so I doubt that it is different.

", "votes": 0, "replies": [] }, { "id": 471062, "author_name": "Bai", "author_url": "", "post_date": "2019-02-14T01:25:27.527000", "content": "

@philippsinger\nbut we could not know the data used in Private LB.

", "votes": -1, "replies": [] } ] }, { "id": 471552, "author_name": "CPMP", "author_url": "", "post_date": "2019-02-14T15:45:20.783000", "content": "

Well, this is not possible after all...

\n\n

They dropped to 1027 position. I hope it is not because of a silly bug in their code.

", "votes": 1, "replies": [] }, { "id": 471272, "author_name": "Yury Kashnitsky", "author_url": "", "post_date": "2019-02-14T08:29:06.153000", "content": "

Nice trolling, Pavel! :)

", "votes": 1, "replies": [] }, { "id": 471201, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2019-02-14T06:22:06.853000", "content": "

@icemekaveli, the shake up answered your question. You jumped higher than that score from 335 position. Congrats.

", "votes": 1, "replies": [ { "id": 471862, "author_name": "Mekaveli.", "author_url": "", "post_date": "2019-02-15T02:09:19.603000", "content": "

Thanks. You score nice too. Congrats.

", "votes": 1, "replies": [] } ] }, { "id": 466879, "author_name": "Anna Novikova", "author_url": "", "post_date": "2019-02-06T06:11:13.463000", "content": "

I don't think they manually labelled submission file, then result would be higher :) Probably we should wait until stage 2 end.

", "votes": 1, "replies": [] }, { "id": 468313, "author_name": "Hasan YILDIZ", "author_url": "", "post_date": "2019-02-08T16:55:36.953000", "content": "

Wouwww, good prize :)

", "votes": 0, "replies": [] }, { "id": 466508, "author_name": "Murat Korkmaz", "author_url": "", "post_date": "2019-02-05T14:22:57.107000", "content": "

I hope we can learn it after the competition.

", "votes": 0, "replies": [] }, { "id": 466490, "author_name": "Ajay Hayagreeve", "author_url": "", "post_date": "2019-02-05T13:36:47.533000", "content": "

Yeah I came to 0.695 after enormous work. Must be a different architechure. Looking forward to learn from them, if they can share...

", "votes": 0, "replies": [] }, { "id": 466566, "author_name": "", "author_url": "", "post_date": "2019-02-05T16:26:08.987000", "content": "", "votes": -1, "replies": [] }, { "id": 466504, "author_name": "", "author_url": "", "post_date": "2019-02-05T14:01:29.770000", "content": "

Truly impressing, they are stratospheric geese, congrats to the team!\nI guess they found something, maybe simple, that we all missed while we were having fun with Neural Nets.\nI am looking forward to learn from them, I hope they will publish their solution.

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "466538": "Let's wait until the second stage results :)", "466428": "Looks like people from different worlds zzz. I'm so looking forward to learn from them after competition ends.", "466462": "I think guys just play game - they use some external data for score on this stage (to practice and scare us :)), but for second stage they will use different model.", "471028": "Narrator: It wasn't possible.", "466433": "Everybody's curious. However, they do have a right not to tell anybody. Anyway, you won't know until medals are distributed :)", "471258": "You could get all the right answers for stage 1 by using Google search API. Toxic questions are not posted, non-toxic questions are on Quora.", "466863": "I could have gotten 0.99 but I held back. I could have labeled 56k questions in 3 days :)", "466443": "Somebody had shared that using BERT, this score is possible.", "467047": "My guess is you somehow derived the original ML-model quora used for labeling. ", "471552": "Well, this is not possible after all...\n\nThey dropped to 1027 position. I hope it is not because of a silly bug in their code.", "471272": "Nice trolling, Pavel! :)", "471201": "@icemekaveli, the shake up answered your question. You jumped higher than that score from 335 position. Congrats.", "466879": "I don't think they manually labelled submission file, then result would be higher :) Probably we should wait until stage 2 end.", "468313": "Wouwww, good prize :)", "466508": "I hope we can learn it after the competition. ", "466490": "Yeah I came to 0.695 after enormous work. Must be a different architechure. Looking forward to learn from them, if they can share...", "466566": "", "466504": "Truly impressing, they are stratospheric geese, congrats to the team!\nI guess they found something, maybe simple, that we all missed while we were having fun with Neural Nets.\nI am looking forward to learn from them, I hope they will publish their solution." } }