For me,\nval AUC 0.92846, LB 0.9393. single model single fold. \nupdate:\nval AUC 0.93300, LB 0.9410. single model single fold.\nupdate:\nLB 0.9483. single model single fold.
", "messages": [ { "id": 800657, "postDate": "2020-04-07T16:02:11.240Z", "content": "For me,\nval AUC 0.92846, LB 0.9393. single model single fold. \nupdate:\nval AUC 0.93300, LB 0.9410. single model single fold.\nupdate:\nLB 0.9483. single model single fold.
", "rawMarkdown": "For me,\nval AUC 0.92846, LB 0.9393. single model single fold. \nupdate:\nval AUC 0.93300, LB 0.9410. single model single fold.\nupdate:\nLB 0.9483. single model single fold.\n", "votes": 29 }, { "id": 800823, "postDate": "2020-04-07T19:14:34.457Z", "content": "\nsingle model: `distilbert-base-multilingual-cased`\nCV: 0.9228 \nLB: 0.9320\n
Update:
\n\n\nmodel: xlm-roberta\nLB: 0.9449\n
Nice score especially with DistilBERT! Did you train on validation set ?
", "rawMarkdown": "Nice score especially with DistilBERT! Did you train on validation set ?", "votes": 1 }, { "id": 831608, "postDate": "2020-05-03T13:32:37.860Z", "content": "Sorry for late reply, didn't get the notification. :(\nNo, I didn't train on the validation set.
", "rawMarkdown": "Sorry for late reply, didn't get the notification. :(\nNo, I didn't train on the validation set." }, { "id": 840405, "postDate": "2020-05-10T01:49:05.540Z", "content": "Wow, LB 9449 single model? that's amazing man...
", "rawMarkdown": "Wow, LB 9449 single model? that's amazing man..." }, { "id": 863168, "postDate": "2020-05-27T05:53:10.393Z", "content": "@ipythonx Can you give some insights please? Amazing job btw!!
", "rawMarkdown": "@ipythonx Can you give some insights please? Amazing job btw!!\n" } ] }, { "id": 840061, "postDate": "2020-05-09T18:42:22.087Z", "content": "XLM-Roberta-large\nCV 0.9520, LB 0.9430
", "rawMarkdown": "XLM-Roberta-large\nCV 0.9520, LB 0.9430", "votes": 3, "replies": [ { "id": 840071, "postDate": "2020-05-09T18:51:09.047Z", "content": "holy moly! \nAre you using translated or raw text? 👀
", "rawMarkdown": "holy moly! \nAre you using translated or raw text? 👀 ", "votes": 1 } ] }, { "id": 844517, "postDate": "2020-05-12T16:49:51.813Z", "content": "CV 0.9580 and LB 0.9465
", "rawMarkdown": "CV 0.9580 and LB 0.9465", "votes": 4, "replies": [ { "id": 851808, "postDate": "2020-05-18T00:05:46.180Z", "content": "Which model are you using ?! are you used Kfold ?
", "rawMarkdown": "Which model are you using ?! are you used Kfold ? " }, { "id": 853547, "postDate": "2020-05-19T09:19:40.017Z", "content": "not CV maybe :)
", "rawMarkdown": "not CV maybe :)" }, { "id": 858706, "postDate": "2020-05-23T17:58:18.817Z", "content": "I'm pretty sure I can make the difference between CV and validation set :)
", "rawMarkdown": "I'm pretty sure I can make the difference between CV and validation set :)", "votes": 1 }, { "id": 878523, "postDate": "2020-06-08T16:05:17.307Z", "content": "Nice. Have you checked any single fold score and how much kfold brings?
", "rawMarkdown": "Nice. Have you checked any single fold score and how much kfold brings? " } ] }, { "id": 800863, "postDate": "2020-04-07T20:02:35.740Z", "content": "xlm-roberta\nLB: 0.9390
", "rawMarkdown": "xlm-roberta\nLB: 0.9390", "votes": 3 }, { "id": 869625, "postDate": "2020-06-01T06:58:47.510Z", "content": ".9427 Roberta
", "rawMarkdown": ".9427 Roberta", "votes": 1, "replies": [ { "id": 869778, "postDate": "2020-06-01T09:28:25.623Z", "content": "Roberta, really?
", "rawMarkdown": "Roberta, really?" }, { "id": 869906, "postDate": "2020-06-01T11:06:23.853Z", "content": "RoBERTa or XLM-RoBERTa?
", "rawMarkdown": "RoBERTa or XLM-RoBERTa?" } ] }, { "id": 852599, "postDate": "2020-05-18T14:49:17.917Z", "content": "CV 0.9534, LB 0.9426
", "rawMarkdown": "CV 0.9534, LB 0.9426\n", "votes": 1, "replies": [ { "id": 852603, "postDate": "2020-05-18T14:52:23.093Z", "content": "can you explain some insights :)
", "rawMarkdown": "can you explain some insights :) " } ] }, { "id": 852585, "postDate": "2020-05-18T14:39:32.650Z", "content": "XLM-Roberta \nSingle model, single fold, trained on undersampled English data as well as a few epochs on validation data.\nCV 0.9120 and LB 0.9402
\n\nUpdate:\nValidation AUC: 0.9438, LB: 0.9418 (Single Model, single fold)
", "rawMarkdown": "XLM-Roberta \nSingle model, single fold, trained on undersampled English data as well as a few epochs on validation data.\nCV 0.9120 and LB 0.9402\n\nUpdate:\nValidation AUC: 0.9438, LB: 0.9418 (Single Model, single fold)", "votes": 1 }, { "id": 851719, "postDate": "2020-05-17T21:09:25.710Z", "content": "In all of these numbers that you guys are mentioning, do you mean out of fold prediction performance on the raw data? translated data? or the hold out validation set we were provided? Are you guys really doing CV or do you just mean validation? Seems like no one uses the term CV properly anymore.
", "rawMarkdown": "In all of these numbers that you guys are mentioning, do you mean out of fold prediction performance on the raw data? translated data? or the hold out validation set we were provided? Are you guys really doing CV or do you just mean validation? Seems like no one uses the term CV properly anymore. ", "votes": 1 }, { "id": 821224, "postDate": "2020-04-26T02:57:44.010Z", "content": "XLM-Roberta \nbase \nCV 0.933 LB 0.9383 1model 5folds mean on validation\nlarge\nCV 0.933 LB 0.9402
", "rawMarkdown": "XLM-Roberta \nbase \nCV 0.933 LB 0.9383 1model 5folds mean on validation\nlarge\nCV 0.933 LB 0.9402", "votes": 1, "replies": [ { "id": 831286, "postDate": "2020-05-03T09:23:54.423Z", "content": "Nice, are you using translated data @goldenlock
", "rawMarkdown": "Nice, are you using translated data @goldenlock " }, { "id": 831599, "postDate": "2020-05-03T13:26:52.873Z", "content": "@shahules Using translated data, *cleaned.csv help a lot.
", "rawMarkdown": "@shahules Using translated data, *cleaned.csv help a lot.", "votes": 2 }, { "id": 840116, "postDate": "2020-05-09T19:27:04.083Z", "content": "Without 5folds, how much did you get with base model ? I always had results below 0.93 on LB
", "rawMarkdown": "Without 5folds, how much did you get with base model ? I always had results below 0.93 on LB " }, { "id": 843643, "postDate": "2020-05-12T07:16:06.003Z", "content": "Oh I did not try, since only 5folds on validation data, for base model it is not costing much.
", "rawMarkdown": "Oh I did not try, since only 5folds on validation data, for base model it is not costing much.", "votes": 1 } ] }, { "id": 807660, "postDate": "2020-04-14T19:23:57.830Z", "content": "Thanks for all your hard work, but can you add your methods too or any useful ideas for beginners to learn. And thanks a lot again. \nBests,
", "rawMarkdown": "Thanks for all your hard work, but can you add your methods too or any useful ideas for beginners to learn. And thanks a lot again. \nBests,", "votes": 1 }, { "id": 802655, "postDate": "2020-04-09T17:30:28.140Z", "content": "For those scores reported, all of you trained with validation dataset?
", "rawMarkdown": "For those scores reported, all of you trained with validation dataset?", "votes": 1, "replies": [ { "id": 803415, "postDate": "2020-04-10T13:52:43.950Z", "content": "As the title saying \"BEST\". I think they will train with validation data if it helps. And whether valid data need to be trained is determined mostly by what kind of model you choose.
", "rawMarkdown": "As the title saying \"BEST\". I think they will train with validation data if it helps. And whether valid data need to be trained is determined mostly by what kind of model you choose.", "votes": 1 } ] }, { "id": 817737, "postDate": "2020-04-23T11:38:03.760Z", "content": "I tried to test approach using large and base model : \ntrain on english and validation (split 90/10) :
\n\nlarge AUC VAL : 0.962 / AUC TEST 0.9407\nbase AUC val : 0.961 / AUC Test 0.924
\n\nSeems that base model are less robust the the large
", "rawMarkdown": "I tried to test approach using large and base model : \ntrain on english and validation (split 90/10) : \n\nlarge AUC VAL : 0.962 / AUC TEST 0.9407\nbase AUC val : 0.961 / AUC Test 0.924\n\nSeems that base model are less robust the the large", "votes": 2 }, { "id": 805992, "postDate": "2020-04-13T10:36:17.290Z", "content": "XLM-Roberta
\n\nVal score: 0.950281\nLB: 0.9405
", "rawMarkdown": "XLM-Roberta\n\nVal score: 0.950281\nLB: 0.9405", "votes": 2, "replies": [ { "id": 806091, "postDate": "2020-04-13T12:40:45.843Z", "content": "Great! did you use validation set for training as well?
", "rawMarkdown": "Great! did you use validation set for training as well?", "votes": 1 }, { "id": 806108, "postDate": "2020-04-13T12:53:00.440Z", "content": "Same question.
", "rawMarkdown": "Same question. ", "votes": 1 }, { "id": 807199, "postDate": "2020-04-14T13:22:09.613Z", "content": "Yes indeed. It gives me the extra boost to pass 0.94
", "rawMarkdown": "Yes indeed. It gives me the extra boost to pass 0.94" }, { "id": 807244, "postDate": "2020-04-14T14:00:07.043Z", "content": "@rftexas Thanks! If so, what is your cv strategy? Split validation set to 5 fold or something?
", "rawMarkdown": "@rftexas Thanks! If so, what is your cv strategy? Split validation set to 5 fold or something?" }, { "id": 808526, "postDate": "2020-04-15T13:03:39.957Z", "content": "Exactly!
", "rawMarkdown": "Exactly!" }, { "id": 808677, "postDate": "2020-04-15T14:55:30.890Z", "content": "Thanks, I'll try too:)
", "rawMarkdown": "Thanks, I'll try too:)" } ] }, { "id": 802791, "postDate": "2020-04-09T20:23:16.883Z", "content": "Single model, without training on validation data:
\n\nVal score: 0.94491\nLB: 0.9366
", "rawMarkdown": "Single model, without training on validation data:\n\nVal score: 0.94491\nLB: 0.9366", "votes": 2, "replies": [ { "id": 803418, "postDate": "2020-04-10T13:54:10.037Z", "content": "wow. nice training!
", "rawMarkdown": "wow. nice training!", "votes": 2 }, { "id": 803444, "postDate": "2020-04-10T14:37:57Z", "content": "That's really impressive. As you said, without training on validation, so that means no translation!
", "rawMarkdown": "That's really impressive. As you said, without training on validation, so that means no translation! " } ] }, { "id": 800873, "postDate": "2020-04-07T20:19:20.840Z", "content": "XLM-Roberta\nLB: 0.9372
", "rawMarkdown": "XLM-Roberta\nLB: 0.9372", "votes": 2 }, { "id": 863146, "postDate": "2020-05-27T05:20:00.180Z", "content": "@mcggood 9483 using single model single fold?? that is crazy!! Well done!
", "rawMarkdown": "@mcggood 9483 using single model single fold?? that is crazy!! Well done!\n", "replies": [ { "id": 863278, "postDate": "2020-05-27T07:44:36.513Z", "content": "Nice to see your big improvement in 2 hours. good job
", "rawMarkdown": "Nice to see your big improvement in 2 hours. good job" }, { "id": 865601, "postDate": "2020-05-28T18:21:28.207Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 854019, "postDate": "2020-05-19T17:18:22.833Z", "content": "CV 0.951, LB 0.9425
\n\nI would like to add a question 😄 : who manage to easily reproduce the announced score ?\nSame 'code' and seeding numpy, random, torch... and i can get 0.9363\nI tried to seed everything but i must miss something...
", "rawMarkdown": "CV 0.951, LB 0.9425\n\nI would like to add a question 😄 : who manage to easily reproduce the announced score ?\nSame 'code' and seeding numpy, random, torch... and i can get 0.9363\nI tried to seed everything but i must miss something...", "replies": [ { "id": 854203, "postDate": "2020-05-19T20:57:31.527Z", "content": "same issue, but I am using TF. without changing seed results are around 0.938 and 0.942X. If I change the seed of the splitting for CV for instance the results can change a lot ... like dropping to 0.92 . validation.csv is probably too small unfortunately
", "rawMarkdown": "same issue, but I am using TF. without changing seed results are around 0.938 and 0.942X. If I change the seed of the splitting for CV for instance the results can change a lot ... like dropping to 0.92 . validation.csv is probably too small unfortunately" }, { "id": 865115, "postDate": "2020-05-28T11:54:03.917Z", "content": "I managed now to have stable results now by increasing the batch size from 16 to 64 (gradient accumulation).
", "rawMarkdown": "I managed now to have stable results now by increasing the batch size from 16 to 64 (gradient accumulation).", "votes": 1 } ] }, { "id": 818896, "postDate": "2020-04-24T07:37:44.573Z", "content": "single model: xlm-roberta\nsingle fold and train with validation dataset\nLB 0.9410
", "rawMarkdown": "single model: xlm-roberta\nsingle fold and train with validation dataset\nLB 0.9410", "replies": [ { "id": 818949, "postDate": "2020-04-24T08:25:24.640Z", "content": "Did you use translated dataset for training? Or you are only using the datasets provided?
", "rawMarkdown": "Did you use translated dataset for training? Or you are only using the datasets provided?" }, { "id": 819027, "postDate": "2020-04-24T09:44:14.383Z", "content": "@tanulsingh077 Yes, i used translated dataset for training, how about you?
", "rawMarkdown": "@tanulsingh077 Yes, i used translated dataset for training, how about you?" }, { "id": 819146, "postDate": "2020-04-24T11:39:54.023Z", "content": "Yes I am trying to do the same, what about data preprocessing, also are you using translation from all the six langauges?
", "rawMarkdown": "Yes I am trying to do the same, what about data preprocessing, also are you using translation from all the six langauges?" }, { "id": 820033, "postDate": "2020-04-25T05:26:23.540Z", "content": "Used translation from 6 languages is good for me, and I trying translated testing set
", "rawMarkdown": "Used translation from 6 languages is good for me, and I trying translated testing set" }, { "id": 821233, "postDate": "2020-04-26T03:19:38.523Z", "content": "are you using public translated data?
", "rawMarkdown": "are you using public translated data?" }, { "id": 822557, "postDate": "2020-04-27T01:50:26.443Z", "content": "@mcggood Yes, i am using public data from Michael Kazachok
", "rawMarkdown": "@mcggood Yes, i am using public data from [Michael Kazachok](https://www.kaggle.com/miklgr500/jigsaw-train-multilingual-coments-google-api#jigsaw-toxic-comment-train-google-fr-cleaned.csv)\n" } ] }, { "id": 817834, "postDate": "2020-04-23T12:59:30.713Z", "content": "Xlm-Roberta Large\nCV- 0.95\nLB-0.938
", "rawMarkdown": "Xlm-Roberta Large\nCV- 0.95\nLB-0.938" }, { "id": 806703, "postDate": "2020-04-14T01:09:02.353Z", "content": "sigle model, single fold, without using validation set for training\ncv: 0.9453\nlb: 0.9379
", "rawMarkdown": "sigle model, single fold, without using validation set for training\ncv: 0.9453\nlb: 0.9379\n", "replies": [ { "id": 807368, "postDate": "2020-04-14T15:51:23.873Z", "content": "are you using translated data for training?
", "rawMarkdown": "are you using translated data for training?" }, { "id": 807900, "postDate": "2020-04-15T02:45:52.717Z", "content": "@mcggood Yes. How about you?
", "rawMarkdown": "@mcggood Yes. How about you?", "votes": 1 }, { "id": 808248, "postDate": "2020-04-15T08:47:20.910Z", "content": "I tried both use and not use. still testing which is better for me
", "rawMarkdown": "I tried both use and not use. still testing which is better for me" }, { "id": 808514, "postDate": "2020-04-15T12:55:37.983Z", "content": "Really! That means sometimes English text is more useful for you, interesting..\nThanks!
", "rawMarkdown": "Really! That means sometimes English text is more useful for you, interesting..\nThanks!" } ] } ], "comments": [ { "id": 800823, "author_name": "Innat", "author_url": "", "post_date": "2020-04-07T19:14:34.457000", "content": "\nsingle model: `distilbert-base-multilingual-cased`\nCV: 0.9228 \nLB: 0.9320\n
Update:
\n\n\nmodel: xlm-roberta\nLB: 0.9449\n
Nice score especially with DistilBERT! Did you train on validation set ?
", "votes": 1, "replies": [] }, { "id": 831608, "author_name": "Innat", "author_url": "", "post_date": "2020-05-03T13:32:37.860000", "content": "Sorry for late reply, didn't get the notification. :(\nNo, I didn't train on the validation set.
", "votes": 0, "replies": [] }, { "id": 840405, "author_name": "DHZM", "author_url": "", "post_date": "2020-05-10T01:49:05.540000", "content": "Wow, LB 9449 single model? that's amazing man...
", "votes": 0, "replies": [] }, { "id": 863168, "author_name": "Amogh ", "author_url": "", "post_date": "2020-05-27T05:53:10.393000", "content": "@ipythonx Can you give some insights please? Amazing job btw!!
", "votes": 0, "replies": [] } ] }, { "id": 840061, "author_name": "Xuan Cao", "author_url": "", "post_date": "2020-05-09T18:42:22.087000", "content": "XLM-Roberta-large\nCV 0.9520, LB 0.9430
", "votes": 3, "replies": [ { "id": 840071, "author_name": "Innat", "author_url": "", "post_date": "2020-05-09T18:51:09.047000", "content": "holy moly! \nAre you using translated or raw text? 👀
", "votes": 1, "replies": [] } ] }, { "id": 844517, "author_name": "Serigne ", "author_url": "", "post_date": "2020-05-12T16:49:51.813000", "content": "CV 0.9580 and LB 0.9465
", "votes": 4, "replies": [ { "id": 851808, "author_name": "Ashraf Mahdhi", "author_url": "", "post_date": "2020-05-18T00:05:46.180000", "content": "Which model are you using ?! are you used Kfold ?
", "votes": 0, "replies": [] }, { "id": 853547, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-05-19T09:19:40.017000", "content": "not CV maybe :)
", "votes": 0, "replies": [] }, { "id": 858706, "author_name": "Serigne ", "author_url": "", "post_date": "2020-05-23T17:58:18.817000", "content": "I'm pretty sure I can make the difference between CV and validation set :)
", "votes": 1, "replies": [] }, { "id": 878523, "author_name": "Miroslav Valan", "author_url": "", "post_date": "2020-06-08T16:05:17.307000", "content": "Nice. Have you checked any single fold score and how much kfold brings?
", "votes": 0, "replies": [] } ] }, { "id": 800863, "author_name": "ma7555", "author_url": "", "post_date": "2020-04-07T20:02:35.740000", "content": "xlm-roberta\nLB: 0.9390
", "votes": 3, "replies": [] }, { "id": 869625, "author_name": "Ashish Gupta", "author_url": "", "post_date": "2020-06-01T06:58:47.510000", "content": ".9427 Roberta
", "votes": 1, "replies": [ { "id": 869778, "author_name": "Psi", "author_url": "", "post_date": "2020-06-01T09:28:25.623000", "content": "Roberta, really?
", "votes": 0, "replies": [] }, { "id": 869906, "author_name": "Ilham Firdausi Putra", "author_url": "", "post_date": "2020-06-01T11:06:23.853000", "content": "RoBERTa or XLM-RoBERTa?
", "votes": 0, "replies": [] } ] }, { "id": 852599, "author_name": "PAB97", "author_url": "", "post_date": "2020-05-18T14:49:17.917000", "content": "CV 0.9534, LB 0.9426
", "votes": 1, "replies": [ { "id": 852603, "author_name": "Ashraf Mahdhi", "author_url": "", "post_date": "2020-05-18T14:52:23.093000", "content": "can you explain some insights :)
", "votes": 0, "replies": [] } ] }, { "id": 852585, "author_name": "Alan Sun", "author_url": "", "post_date": "2020-05-18T14:39:32.650000", "content": "XLM-Roberta \nSingle model, single fold, trained on undersampled English data as well as a few epochs on validation data.\nCV 0.9120 and LB 0.9402
\n\nUpdate:\nValidation AUC: 0.9438, LB: 0.9418 (Single Model, single fold)
", "votes": 1, "replies": [] }, { "id": 851719, "author_name": "ryches", "author_url": "", "post_date": "2020-05-17T21:09:25.710000", "content": "In all of these numbers that you guys are mentioning, do you mean out of fold prediction performance on the raw data? translated data? or the hold out validation set we were provided? Are you guys really doing CV or do you just mean validation? Seems like no one uses the term CV properly anymore.
", "votes": 1, "replies": [] }, { "id": 821224, "author_name": "gezi", "author_url": "", "post_date": "2020-04-26T02:57:44.010000", "content": "XLM-Roberta \nbase \nCV 0.933 LB 0.9383 1model 5folds mean on validation\nlarge\nCV 0.933 LB 0.9402
", "votes": 1, "replies": [ { "id": 831286, "author_name": "Shahules", "author_url": "", "post_date": "2020-05-03T09:23:54.423000", "content": "Nice, are you using translated data @goldenlock
", "votes": 0, "replies": [] }, { "id": 831599, "author_name": "gezi", "author_url": "", "post_date": "2020-05-03T13:26:52.873000", "content": "@shahules Using translated data, *cleaned.csv help a lot.
", "votes": 2, "replies": [] }, { "id": 840116, "author_name": "Shiro", "author_url": "", "post_date": "2020-05-09T19:27:04.083000", "content": "Without 5folds, how much did you get with base model ? I always had results below 0.93 on LB
", "votes": 0, "replies": [] }, { "id": 843643, "author_name": "gezi", "author_url": "", "post_date": "2020-05-12T07:16:06.003000", "content": "Oh I did not try, since only 5folds on validation data, for base model it is not costing much.
", "votes": 1, "replies": [] } ] }, { "id": 807660, "author_name": "Mohammed Deeb", "author_url": "", "post_date": "2020-04-14T19:23:57.830000", "content": "Thanks for all your hard work, but can you add your methods too or any useful ideas for beginners to learn. And thanks a lot again. \nBests,
", "votes": 1, "replies": [] }, { "id": 802655, "author_name": "Yih-Dar SHIEH", "author_url": "", "post_date": "2020-04-09T17:30:28.140000", "content": "For those scores reported, all of you trained with validation dataset?
", "votes": 1, "replies": [ { "id": 803415, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-04-10T13:52:43.950000", "content": "As the title saying \"BEST\". I think they will train with validation data if it helps. And whether valid data need to be trained is determined mostly by what kind of model you choose.
", "votes": 1, "replies": [] } ] }, { "id": 817737, "author_name": "Shiro", "author_url": "", "post_date": "2020-04-23T11:38:03.760000", "content": "I tried to test approach using large and base model : \ntrain on english and validation (split 90/10) :
\n\nlarge AUC VAL : 0.962 / AUC TEST 0.9407\nbase AUC val : 0.961 / AUC Test 0.924
\n\nSeems that base model are less robust the the large
", "votes": 2, "replies": [] }, { "id": 805992, "author_name": "PAB97", "author_url": "", "post_date": "2020-04-13T10:36:17.290000", "content": "XLM-Roberta
\n\nVal score: 0.950281\nLB: 0.9405
", "votes": 2, "replies": [ { "id": 806091, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-13T12:40:45.843000", "content": "Great! did you use validation set for training as well?
", "votes": 1, "replies": [] }, { "id": 806108, "author_name": "Innat", "author_url": "", "post_date": "2020-04-13T12:53:00.440000", "content": "Same question.
", "votes": 1, "replies": [] }, { "id": 807199, "author_name": "PAB97", "author_url": "", "post_date": "2020-04-14T13:22:09.613000", "content": "Yes indeed. It gives me the extra boost to pass 0.94
", "votes": 0, "replies": [] }, { "id": 807244, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-14T14:00:07.043000", "content": "@rftexas Thanks! If so, what is your cv strategy? Split validation set to 5 fold or something?
", "votes": 0, "replies": [] }, { "id": 808526, "author_name": "PAB97", "author_url": "", "post_date": "2020-04-15T13:03:39.957000", "content": "Exactly!
", "votes": 0, "replies": [] }, { "id": 808677, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-15T14:55:30.890000", "content": "Thanks, I'll try too:)
", "votes": 0, "replies": [] } ] }, { "id": 802791, "author_name": "Alex Shonenkov", "author_url": "", "post_date": "2020-04-09T20:23:16.883000", "content": "Single model, without training on validation data:
\n\nVal score: 0.94491\nLB: 0.9366
", "votes": 2, "replies": [ { "id": 803418, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-04-10T13:54:10.037000", "content": "wow. nice training!
", "votes": 2, "replies": [] }, { "id": 803444, "author_name": "Innat", "author_url": "", "post_date": "2020-04-10T14:37:57", "content": "That's really impressive. As you said, without training on validation, so that means no translation!
", "votes": 0, "replies": [] } ] }, { "id": 800873, "author_name": "PAB97", "author_url": "", "post_date": "2020-04-07T20:19:20.840000", "content": "XLM-Roberta\nLB: 0.9372
", "votes": 2, "replies": [] }, { "id": 863146, "author_name": "Amogh ", "author_url": "", "post_date": "2020-05-27T05:20:00.180000", "content": "@mcggood 9483 using single model single fold?? that is crazy!! Well done!
", "votes": 0, "replies": [ { "id": 863278, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-05-27T07:44:36.513000", "content": "Nice to see your big improvement in 2 hours. good job
", "votes": 0, "replies": [] }, { "id": 865601, "author_name": "", "author_url": "", "post_date": "2020-05-28T18:21:28.207000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 854019, "author_name": "AnhMeow", "author_url": "", "post_date": "2020-05-19T17:18:22.833000", "content": "CV 0.951, LB 0.9425
\n\nI would like to add a question 😄 : who manage to easily reproduce the announced score ?\nSame 'code' and seeding numpy, random, torch... and i can get 0.9363\nI tried to seed everything but i must miss something...
", "votes": 0, "replies": [ { "id": 854203, "author_name": "Shiro", "author_url": "", "post_date": "2020-05-19T20:57:31.527000", "content": "same issue, but I am using TF. without changing seed results are around 0.938 and 0.942X. If I change the seed of the splitting for CV for instance the results can change a lot ... like dropping to 0.92 . validation.csv is probably too small unfortunately
", "votes": 0, "replies": [] }, { "id": 865115, "author_name": "AnhMeow", "author_url": "", "post_date": "2020-05-28T11:54:03.917000", "content": "I managed now to have stable results now by increasing the batch size from 16 to 64 (gradient accumulation).
", "votes": 1, "replies": [] } ] }, { "id": 818896, "author_name": "Arctic Yen", "author_url": "", "post_date": "2020-04-24T07:37:44.573000", "content": "single model: xlm-roberta\nsingle fold and train with validation dataset\nLB 0.9410
", "votes": 0, "replies": [ { "id": 818949, "author_name": "Mr_KnowNothing", "author_url": "", "post_date": "2020-04-24T08:25:24.640000", "content": "Did you use translated dataset for training? Or you are only using the datasets provided?
", "votes": 0, "replies": [] }, { "id": 819027, "author_name": "Arctic Yen", "author_url": "", "post_date": "2020-04-24T09:44:14.383000", "content": "@tanulsingh077 Yes, i used translated dataset for training, how about you?
", "votes": 0, "replies": [] }, { "id": 819146, "author_name": "Mr_KnowNothing", "author_url": "", "post_date": "2020-04-24T11:39:54.023000", "content": "Yes I am trying to do the same, what about data preprocessing, also are you using translation from all the six langauges?
", "votes": 0, "replies": [] }, { "id": 820033, "author_name": "Arctic Yen", "author_url": "", "post_date": "2020-04-25T05:26:23.540000", "content": "Used translation from 6 languages is good for me, and I trying translated testing set
", "votes": 0, "replies": [] }, { "id": 821233, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-04-26T03:19:38.523000", "content": "are you using public translated data?
", "votes": 0, "replies": [] }, { "id": 822557, "author_name": "Arctic Yen", "author_url": "", "post_date": "2020-04-27T01:50:26.443000", "content": "@mcggood Yes, i am using public data from Michael Kazachok
", "votes": 0, "replies": [] } ] }, { "id": 817834, "author_name": "Mr_KnowNothing", "author_url": "", "post_date": "2020-04-23T12:59:30.713000", "content": "Xlm-Roberta Large\nCV- 0.95\nLB-0.938
", "votes": 0, "replies": [] }, { "id": 806703, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-14T01:09:02.353000", "content": "sigle model, single fold, without using validation set for training\ncv: 0.9453\nlb: 0.9379
", "votes": 0, "replies": [ { "id": 807368, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-04-14T15:51:23.873000", "content": "are you using translated data for training?
", "votes": 0, "replies": [] }, { "id": 807900, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-15T02:45:52.717000", "content": "@mcggood Yes. How about you?
", "votes": 1, "replies": [] }, { "id": 808248, "author_name": "MaChaogong", "author_url": "", "post_date": "2020-04-15T08:47:20.910000", "content": "I tried both use and not use. still testing which is better for me
", "votes": 0, "replies": [] }, { "id": 808514, "author_name": "Camaro", "author_url": "", "post_date": "2020-04-15T12:55:37.983000", "content": "Really! That means sometimes English text is more useful for you, interesting..\nThanks!
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "800657": "For me,\nval AUC 0.92846, LB 0.9393. single model single fold. \nupdate:\nval AUC 0.93300, LB 0.9410. single model single fold.\nupdate:\nLB 0.9483. single model single fold.\n", "800823": "```\nsingle model: `distilbert-base-multilingual-cased`\nCV: 0.9228 \nLB: 0.9320\n```\n\nUpdate:\n\n```\nmodel: xlm-roberta\nLB: 0.9449\n```\n", "840061": "XLM-Roberta-large\nCV 0.9520, LB 0.9430", "844517": "CV 0.9580 and LB 0.9465", "800863": "xlm-roberta\nLB: 0.9390", "869625": ".9427 Roberta", "852599": "CV 0.9534, LB 0.9426\n", "852585": "XLM-Roberta \nSingle model, single fold, trained on undersampled English data as well as a few epochs on validation data.\nCV 0.9120 and LB 0.9402\n\nUpdate:\nValidation AUC: 0.9438, LB: 0.9418 (Single Model, single fold)", "851719": "In all of these numbers that you guys are mentioning, do you mean out of fold prediction performance on the raw data? translated data? or the hold out validation set we were provided? Are you guys really doing CV or do you just mean validation? Seems like no one uses the term CV properly anymore. ", "821224": "XLM-Roberta \nbase \nCV 0.933 LB 0.9383 1model 5folds mean on validation\nlarge\nCV 0.933 LB 0.9402", "807660": "Thanks for all your hard work, but can you add your methods too or any useful ideas for beginners to learn. And thanks a lot again. \nBests,", "802655": "For those scores reported, all of you trained with validation dataset?", "817737": "I tried to test approach using large and base model : \ntrain on english and validation (split 90/10) : \n\nlarge AUC VAL : 0.962 / AUC TEST 0.9407\nbase AUC val : 0.961 / AUC Test 0.924\n\nSeems that base model are less robust the the large", "805992": "XLM-Roberta\n\nVal score: 0.950281\nLB: 0.9405", "802791": "Single model, without training on validation data:\n\nVal score: 0.94491\nLB: 0.9366", "800873": "XLM-Roberta\nLB: 0.9372", "863146": "@mcggood 9483 using single model single fold?? that is crazy!! Well done!\n", "854019": "CV 0.951, LB 0.9425\n\nI would like to add a question 😄 : who manage to easily reproduce the announced score ?\nSame 'code' and seeding numpy, random, torch... and i can get 0.9363\nI tried to seed everything but i must miss something...", "818896": "single model: xlm-roberta\nsingle fold and train with validation dataset\nLB 0.9410", "817834": "Xlm-Roberta Large\nCV- 0.95\nLB-0.938", "806703": "sigle model, single fold, without using validation set for training\ncv: 0.9453\nlb: 0.9379\n" } }