Datasets:

Abirate
/

all_write_ups

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 3.85k rows

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.

Title of Competition stringclasses 343 values	Title of Writeup stringlengths 6 139	User stringclasses 1 value	Writeup stringlengths 419 80.4k
Jigsaw Multilingual Toxic Comment Classification	8th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Multilingual Toxic Comment Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I tried a few sampling techniques and as said I dont think it has any huge impact. I rather did some downsampling in the end to speed up training.</p> <p>In the end only using public LB.</p> <p>Thanks a lot to Kaggle and Jigsaw to host this competition. I am very happy to achieve my second solo gold medal within a week :)</p> <p>My main goal when starting this competition was to learn how to use TPUs on Kaggle. I started with Pytorch, switched to Tensorflow, and then switched back to Pytorch. I think overall the usability is still very shaky, and 3 hour limit is also limiting things quite a bit. There is constant progress on it though, and I hope that they will be better usable in the future. I made a separate <a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/159723">topic</a> before.</p> <p>The whole test set was visible in this competition, so subbing was very easy and blending many models was very important. Unfortunately, I could not select my best sub because it included a blend with a <a href="https://www.kaggle.com/shonenkov/tpu-inference-super-fast-xlmroberta">kernel</a> where no training script was available. I am happy I followed my heart to not select it even though I would have ranked a few spots higher with it. I really would encourage people to stop posting inference only kernels, because logically people will use them if they score well, but the learning potential is quite limited imho.</p> <h3>Models</h3> <p>There is not much differences in the models I fit to those that have been posted throughout the competition. I could not get any better results compared to using XLM-Roberta-Large. I also tried uni-language models without any improvements over XLM-Roberta.</p> <p>I also do not believe that any advanced sampling technique, label smoothing, augmentation, or other stuff floating around was giving any improvements. </p> <p>The only real important thing was to translate the training data to the languages at hand. This is in line with the experiments reported in the original <a href="https://arxiv.org/pdf/1911.02116.pdf">XLM-Roberta paper</a>.</p> <p>I blended models by doing rank-averaged blending.</p> <h3>Data</h3> <p>I only used provided data and translations of it. I think the most untapped potential was to use talk pages from Wikipedia dumps to further train the models. I did not have time to explore this, but I hope someone else did and can report on it.</p> <p>I actually think one of the main issues of the provided English data was that it included barely any Bot comments, but the test data contained quite many of them, specifically Turkish which is also why I think the score is better there than for others.</p> <h3>Post-processing</h3> <p>After blending, post-processing was quite important for me to reach high scores. As the metric is a global AUC over six different languages, it became quite quickly obvious to me that there should be room to optimize the global score by shifting predictions of individual languages up or down. </p> <p>I tested this on CV with the three languages and managed to get very good boosts there. I further tested how a simulated public and private dataset would behave in terms of optimization, assuming a random split between public and private. So what I did was to repeatedly select random 30% of CV, optimize the thresholds for the 30% and apply them to the remaining 70% and check how far off they would be from the optimal thresholds of the 70%. Overall, I saw that there was not much difference, which is why I was quite confident that I can optimize the thresholds on public LB and they should work well for private LB. Optimizing on CV alone does not work because CV only has 3 languages. This is an example plot of such an analysis:</p> <p><img src="https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F37166%2F14ef07f17e3b1eda7b98ebac1f6cb46b%2Fdownload.png?generation=1592897946838419&alt=media" alt=""></p> <p>The x-axis is the scaling factor for the language at hand, the y-axis is the improvement of the score over no scaling. The blue line is the improvement of the 30% sample and the orange of the 70% sample. Note that the improvement is not 0 for factor 1.0 in this case as another language has already been optimized.</p> <p>Additionally, I optimized a range of values for which I optimize the scaling factors. I found that not scaling the top predictions improved scores, so I only scaled predictions between 0.25 and 0.875 (rank average values). You can get some further minor improvements by averaging the predictions for similar texts.</p> <p>My whole PP routine boosted scores by roughly 0.004, so quite a lot.</p>
Jigsaw Multilingual Toxic Comment Classification	Second silver in one week	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Multilingual Toxic Comment Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks bro and congratulations for 4X Expert !</p> <p>Congratulations all, I am happy to belong to such an amazing community and happy to see new expert like me,new Master like my teammate <a href="/futureboykid">@futureboykid</a> and new GM. see you soon in another leaderboard !</p> <p>Thank you Mehdi ! Congrats to you too </p> <p>Thanks a lot </p>
Jigsaw Multilingual Toxic Comment Classification	🏅 Best single models 🏅	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Multilingual Toxic Comment Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p><a href="/konohayui">@konohayui</a> you can see it at the last section of the model <a href="https://www.kaggle.com/hmendonca/jigsaw20-xlm-r-lb0-9487-singel-model#Apply-language-multipliers">https://www.kaggle.com/hmendonca/jigsaw20-xlm-r-lb0-9487-singel-model#Apply-language-multipliers</a> </p> <p>thanks for let us know <a href="/leecming">@leecming</a> your <a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/160862">write-up</a> is very inspiring! I assume from your msg that you haven't used your distilled model in your final ensemble, is that right? Moreover, how do you prevent your use of PL to interfere with the ensemble diversification?</p> <p>Indeed, I agree you that the vocab embeddings seem to be the bottle neck of the current models, both in learning capacity as well as in number of parameters, which hits me as a possible improvement to the NLP models. However, that might not be a major issue on the new huge models like GPT3, we'll see... My team mate <a href="/ratthachat">@ratthachat</a> and I tried to mitigate the issue adding more layers at the top of the model instead, but only with some moderate success.</p> <p>Thanks <a href="/drpatrickchan">@drpatrickchan</a> Congrats on your first gold medal and the Comp Master! :)</p> <p>I wonder what were the best singles models around.</p> <p>From the write ups it seems like most people got XLM-R around 0.942* and heavily rely on ensembling, post-processing and other tricks (not diminishing their amazing achievement).</p> <p>Anyway, I'd like to share a <a href="https://www.kaggle.com/hmendonca/jigsaw20-xlm-r-lb0-9487-singel-model">single model</a> (weights and code) with pseudo labels and knowledge distillation that got to <strong>LB 0.9475</strong> (and <strong>0.9487</strong> after applying the hardcoded multipliers from @christofhenkel thx!) here <a href="https://www.kaggle.com/hmendonca/jigsaw20-xlm-r-lb0-9487-singel-model">https://www.kaggle.com/hmendonca/jigsaw20-xlm-r-lb0-9487-singel-model</a></p> <p>Unfortunately we failed to ensemble as well as others and did not explore the language modifiers beforehand. This comp was an amazing learning experience for me doing NLP for the first time in my life :D</p> <p>The main take out for me was that relying to heavily on pseudo labels seem to have made blending a much harder task (knowledge distillation seem to have worsen that issue). It took me too long to realize that but lesson learned for the next comps. I'll extend my report when I find some more time, but please let me know your thoughts on that! (and please share your models if you can)</p> <p>Cheers!</p>
Jigsaw Rate Severity of Toxic Comments	118th, Comparison-based Approach	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>In this competition, I used a comparison-based approach. I thought training a comparer was the easiest way to combine all existing datasets: datasets of the two previous competitions, and the Ruddit dataset.</p> <h3>Model</h3> <p>My approach is a little like the popular <code>MarginRankingLoss</code> approach. The <code>MarginRankingLoss</code> compares two output logits to calculate the loss, while I predict logits based on two embeddings and train the models with BCE loss. As shown in the figure below, given two texts, the model needs to predict whether Text 1 is more toxic than Text 2, and the label is 0 or 1.</p> <p>After encoding two comments, I also made a cross-attention to get better representations. You can refer to my inference <a href="https://www.kaggle.com/wuyhbb/deberta-jigsaw-inference" target="_blank">notebook</a> for the detail of the model.</p> <p><img src="https://imgur.com/fXHTSO0" alt="Model Structure"></p> <h3>Data</h3> <p>Jigsaw 1: use all the comments with the label “Toxic” and without the label of “Severe Toxic”, and build random pairs with other comments<br> Jigsaw 2: used comments whose toxicity_annotator_count is large than 10. Randomly build pairs and determined the label with scores calculated as severe_toxicity*8+toxic<br> Ruddit: Randomly build pairs and determined the label with scores.</p> <p>I trained the model with all the data from different sources, and I did not use the validation data to train.</p> <h3>Results</h3> <table> <thead> <tr> <th>Model</th> <th>Validation</th> <th>Private</th> <th>Public</th> </tr> </thead> <tbody> <tr> <td>Roberta Base</td> <td>0.7034</td> <td>0.79937</td> <td>0.74746</td> </tr> <tr> <td>DEberta Base</td> <td>0.7036</td> <td>0.80251</td> <td>0.73669</td> </tr> </tbody> </table> <p>It is a pity that I had a severe bug, and I did not submit the right result.</p> <p>Special Thanks to Z by HP & NVIDIA for sponsoring me a Z4 Workstation with dual A6000 GPU!</p> <p>Training large transformer models with 48GB VRAM is awesome!</p>
Jigsaw Rate Severity of Toxic Comments	11th & 44th place solution - toxidity as percentile	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Always glad to help. :-) <br> On Github we have an example notebooks folder with plenty of stuff.</p> <p>Hi,</p> <p>I am Thomas and creator of an automl library called e2eml (we are still in baby shoes yet).<br> I justed wanted to test the library and see how far I can go with it.</p> <p>I finished 44th place, but also had an unselected submission, which would have hit 11th place. I did not select it, because I thought it would run too long and exceed the 9 hours limit (my fault misassuming that).</p> <p>So what has been the strategy? <br> I started very late into the competition and quickly looked around some notebooks. I found these toxidity weights everyone seemed to use. I did not train my models directly on this because these values seemed a little static. So I followed the following approach:</p> <ul> <li>in a few hunded permutations I sampled from the full dataset</li> <li>based on the toxiidity score I calculated the percentile within that sample (so interpreting toxidity as a relative position)</li> <li>after running all permutations I averaged the percentiles for each comment</li> </ul> <p>From a model architecture the 11th place solution has been a refined Roberta model, which has been pretrained on toxic comments (available on huggingface). I created five model weights on different random splits and averaged them. As it looked too slow on inference I did not chose it for submission.</p> <p>The 44th place solution took only three of these refined Roberta weights and blended it with a SGD on TF-IDF model. </p> <p>Some notebooks including the 44th place inference have been published as examples on our <a href="https://github.com/ThomasMeissnerDS/e2e_ml/tree/develop/Example%20notebooks/Kaggle%20competition%20examples/Jigsaw%20Rate%20Severity%20of%20Toxic%20Comments%202021" target="_blank">Github repo</a>.</p>
Jigsaw Rate Severity of Toxic Comments	12th Solution – Trust Your CV	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h5>First of all, thank you for the organizers and other participants. I’ve learned many things including the importance of CV.</h5> <h4>Solution</h4> <p>My solution is summarized as</p> <ul> <li>Use ‘validation_data.csv’ for the CV estimation. (Do not use it for the training)</li> <li>Training dataset - ‘jigsaw-unintended-bias-training.csv’ and ‘jigsaw-toxic-comment-train.csv’ in ‘<a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification" target="_blank">Jigsaw Multilingual Toxic Comment Classification</a>’ challenge</li> <li>Model – TF-IDF with/wo text cleaning, distilroberta-base, roberta-base, roberta-large</li> <li>Ensemble of all trained model and TF-IDF</li> </ul> <h4>CV</h4> <p>Among my solution, the most important thing is to believe the CV. So, I firstly introduce the discussion that encourages me to focus on CV not LB.</p> <ul> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/287033" target="_blank">Carefull, LB is only 5%</a></li> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/287147" target="_blank">Classic CV vs LB Discussion</a></li> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/289089" target="_blank">Higher CV, lower LB</a></li> </ul> <p>I find that it is hard to obtain CV score over 0.70 (using validation_data.csv), so my goal is to obtain <strong>ensemble of model that stably gives</strong> CV 0.7.</p> <h4>Dataset</h4> <p>It seems most of the public kernel use ’jigsaw-toxic-comment-train.csv’ from ‘Toxic Comment Classification Challenge’. It seems that data is subset of the data in ‘<a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification" target="_blank">Jigsaw Multilingual Toxic Comment Classification</a>’. so I use the data from Multilingual challenge.<br> In the toxic comment and unintended bias, there are columns including ‘toxic’, ‘severe_toxic’, etc. I just give weight 1 for all cols except ‘severe_toxic’ that has weight 2. (It seems there is no huge difference on CV) </p> <h4>Model and Ensemble</h4> <p><strong>TF-IDF</strong> – the code for TF-IDF is barely modified from the referenced code</p> <ul> <li><a href="https://www.kaggle.com/yamqwe/ridge-fasttext-improved-roberta#Toxic-clean-data" target="_blank">🔥 Ridge + fastText + Improved RoBERTa 🔥</a></li> <li><a href="https://www.kaggle.com/vitaleey/tfidf-ridge" target="_blank">TFIDF+RIDGE</a></li> </ul> <p><strong>Bert-like model</strong> - distilroberta-base, roberta-base, and roberta-large. Trained with fasthug. </p> <ul> <li><a href="https://www.kaggle.com/alibaba19/jigsaw-training-ulmfit-w-fastai" target="_blank">Jigsaw Training - ULMFIT w/FastAi</a></li> </ul> <p><strong>Submission (ensemble)</strong> – use exhaustive search and equal weights for the submissions. Both works but equal weights is better.</p> <ul> <li>Sub 1 (CV 0.706, Public 0.77805, Private 0.81029) - exhaustive search to maximizes CV </li> <li>Sub 2 (CV 0.705, Public 0.77631, Private 0.81044) - Similar weights except roberta-large (Actually, in case of roberta-large, I trained few epochs due to the computational budget…)  CV 705, Public 0.77631 Private 0.81044</li> </ul> <p>The inference kernel is updated <a href="https://www.kaggle.com/learnitanyway/12th-solution-jrstc-tf-idf-and-fastai" target="_blank">here</a>. Thank you again for all participants and the organizers and congratulations to all winners.</p>
Jigsaw Rate Severity of Toxic Comments	14th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks and congrats to you too! What would you like to know about the fine-tuning? I think I found that ruddit was less effective than jigsaw 1 or 2, but it did seem to have a positive effect in at least some cases in my experiments.</p> <h1>Overview</h1> <p>The public LB didn't seem very useful so my strategy was to just maximise my validation score. My final submission is a weighted mean of 6 (5-fold) transformers. It was both my highest CV and highest private LB score, so I'm glad I trusted my CV this time.</p> <h1>Data</h1> <p>I tried to include as many different datasets as I could. For training I used:</p> <ul> <li>the validation set</li> <li>jigsaw 1: the data from <a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge" target="_blank">the first jigsaw competition</a></li> <li>jigsaw 2: the data from <a href="https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification" target="_blank">the second jigsaw competition</a></li> <li>ruddit: the data from <a href="https://aclanthology.org/2021.acl-long.210/" target="_blank">Ruddit: Norms of Offensiveness for English Reddit Comments</a></li> <li>offenseval2020: the data from <a href="https://sites.google.com/site/offensevalsharedtask/results-and-paper-submission" target="_blank">OffensEval 2020: Multilingual Offensive Language Identification in Social Media</a>. This dataset was uploaded to haggle by <a href="https://www.kaggle.com/vaby667" target="_blank">@vaby667</a> <a href="https://www.kaggle.com/vaby667/toxictask" target="_blank">here</a></li> </ul> <h3>CV strategy</h3> <p>I used the <a href="https://www.kaggle.com/columbia2131/jigsaw-cv-strategy-by-union-find" target="_blank">Union-Find</a> method by <a href="https://www.kaggle.com/columbia2131" target="_blank">@columbia2131</a> to generate folds that didn't have any texts leaked across folds. I didn't use majority voting or any other method of removing disagreements from the data, as I thought this would just make the validation set artificially easier and less similar to the test data.</p> <h1>Training</h1> <h3>Loss</h3> <p>I used Margin Ranking Loss to train with the validation data, and tried both Margin Ranking and MSE with the other datasets. It was fairly easy to create large amounts of ranked paired data from the extra datasets, but I didn't find that this improved performance over just training them directly on the labels with MSE loss, and also required lower batch sizes.</p> <h3>Evaluation</h3> <p>When fine-tuning on the extra datasets, I computed average agreement with annotators on the validation set at each evaluation, and used early stopping. I trained for multiple epochs on the small datasets (jigsaw 1 and ruddit), evaluating once an epoch. For the large datasets (offenseval and jigsaw 2) I usually only trained for 1 epoch, evaluating every 10% the epoch's steps.</p> <p>When training on the validation data, I trained 5 models, using 4/5 folds for training and the remaining fold as validation for each one. I computed the CV at each epoch and used the model weights from the epoch that had the highest CV. </p> <p>I had original started out using fold-wise early stopping, but I discovered that <a href="https://twitter.com/a_erdem4/status/1483379758331269123" target="_blank">this leads to overly optimistic CV scores</a>.</p> <h3>Multi-stage fine-tuning</h3> <p>I found that taking a model I had already fine-tuned, and fine-tuning it on another dataset improved performance. This worked across multiple fine-tuning stages. The order that worked best was to start by fine-tuning on the larger, and lower scoring, datasets first, and then on the smaller ones after. </p> <p>For example:</p> <ol> <li>fine-tune a pretrained model on offenseval (validation: ~0.68)</li> <li>use #1 to fine-tune on jigsaw 2 (validation ~0.695)</li> <li>fine-tune #2 on jigsaw 1 (validation ~0.7)</li> <li>fine-tune #3 on ruddit (validation ~0.705)</li> <li>fine-tune 5 folds on the validation data, using #4 (CV: 0.71+)</li> </ol> <h3>Hyperparameters</h3> <p>I started out training most of the base-sized models with 1e-5 on earlier fine-tuning stages and reduced to 1-6 on the later ones. I used 1e-6 and then 5e-7 on the larger models. At each stage I also used warmup of 5% and linear LR decay. Even with mixed precision training, I could only fit batch size 8 on the GPU with the large models, so I used gradient accumulation to simulate batch sizes of 64.</p> <h1>Models</h1> <p>Here's a table of model results. </p> <table> <thead> <tr> <th>base model</th> <th>folds</th> <th>CV</th> <th>final submission</th> </tr> </thead> <tbody> <tr> <td>deberta-v3-base</td> <td>5</td> <td>0.715</td> <td>yes</td> </tr> <tr> <td>distilroberta-base</td> <td>5</td> <td>0.714</td> <td>yes</td> </tr> <tr> <td>deberta-v3-large</td> <td>5</td> <td>0.714</td> <td>yes</td> </tr> <tr> <td>deberta-large</td> <td>10</td> <td>0.714</td> <td>yes</td> </tr> <tr> <td>deberta-large</td> <td>5</td> <td>0.713</td> <td>yes</td> </tr> <tr> <td>rembert</td> <td>5</td> <td>0.713</td> <td>yes</td> </tr> <tr> <td>roberta-base</td> <td>5</td> <td>0.711</td> <td>no</td> </tr> <tr> <td>roberta-large</td> <td>5</td> <td>0.708</td> <td>no</td> </tr> </tbody> </table> <ul> <li>I mostly stuck to using roberta and deberta variants because they always perform well. If I had had more time I would've tried some others, but I spent most of the time trying out different combinations of datasets.</li> <li>The reason I tried rembert was because I wanted to make use of the multilingual <a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification" target="_blank">jigsaw 3</a> data. I wasn't able to get any improvement from including the extra data, but I was still able to get reasonably good performance out of rembert.</li> <li>Deberta-large (v1) is in there twice because I did an experiment with a 10 fold model which turned out quite well. I didn't want to keep training 10 fold models though because it took too long.</li> <li>I think all of the large models are slightly under-trained. Training on large datasets like offenseval and jigsaw 2 took over 24 hours so my colab instances timed-out. </li> </ul> <h1>Ensembling</h1> <p>My final submission was a weighted mean of 6 models which were selected from a pool of models by trying to add each of them to the ensemble one at a time, tuning the weights with Optuna for each combination of models, and greedily selecting whichever model increased the OOF score the most. This was repeated until the OOF score stopped improving. My best score was 0.7196.</p> <p>Interestingly, the highest weighted models ended up being deberta-large and rembert, despite those having lower CV scores.</p> <h1>Things which didn't work</h1> <h3>The measuring hate speech dataset</h3> <p>The <a href="https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech" target="_blank">Measuring Hate Speech dataset</a> by ucberkeley-dlab seemed like it was going to be useful, but the labels didn't seem to match the annotations in the validation set for this competition very well. I was unable to get more than 0.656 with this data.</p> <h3>Binary labels</h3> <p>I wanted to try to make use of the binary labelled data too (<a href="https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification" target="_blank">jigsaw 3</a> and <a href="https://www.kaggle.com/ashwiniyer176/toxic-tweets-dataset" target="_blank">toxic tweets</a>). I tried fine-tuning on these datasets, and used the model's predicted probability of the positive class as an output for inference and evaluation. I was able to get 0.68 on the validation set with this method, but I found that it didn't chain together with my multi-stage fine-tuning approach as I had to modify the last layer of the model to switch between regression and classification tasks.</p> <h3>TF-IDF with linear models</h3> <p>I experimented with it a little bit, but wasn't able to get anything over 0.7 on the validation set so I gave up on it.</p> <h3>Word vectors with linear models</h3> <p>I tried encoding each comment as the average of spaCy word vectors, and using this as an input into various linear models. It did about as well as TF-IDF.</p> <h1>Improvements</h1> <p>As I've already mentioned, I think my large models are under-trained due to GPU limitations. I was surprised by how much rembert helped the ensemble, so think I could've made a stronger ensemble by choosing some more diverse model architectures instead of focusing on deberta and roberta so much.</p> <p>I guess the validation scores are not really comparable unless we used the same CV strategy and preprocessing</p> <p>I forgot to add that I made my github repo public too in case anyone wants to see my source code: <a href="https://github.com/AMontgomerie/jigsaw-toxic-severity-competition" target="_blank">https://github.com/AMontgomerie/jigsaw-toxic-severity-competition</a></p> <p>The second submission was just my highest scoring single model: deberta-v3-base 0.715, which got 0.803 on private LB.</p> <p>I tried using hate_speech_score as the target first, which produced the 0.656 score I mentioned. I also tried using the sum of a subset of the labels like ["insult", "humiliate, etc. ] and using this as the target (since this worked well for jigsaw_1) but it performed worse than using the hate_speech_score. </p> <p>I guess there might be some better ways to calculate targets since there are a lot of features in this dataset, but I didn't find a good way to use them.</p> <p>I tried both MSE loss with the original target values and MarginRankingLoss with paired comments and found that MSE worked better. When fine-tuning on ruddit I also found that training for lots of epochs (compared to when fine-tuning on the other datasets) was beneficial.</p>
Jigsaw Rate Severity of Toxic Comments	1st place solution with code	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Apologize for the confusion. Weights were generated during validation (in those valid.py of the training code) by genetic algorithms. You can generate them by re-running my training code (results may differ a little bit due to the randomness of the algorithm). For inference, I simply copied and hard-coded the weights in the script rather than upload them as files (as we all do for transformer weights) because the number of weights are very small.</p> <p>Public LB looks misleading so I focused on the validation performance only. Since it's unclear whether there is any distribution shift between validation data and private LB data, I only considered linear models build on validation data to maximize the robustness of the solution. To achieve this, I trained models on the Jigsaw2018 data and use the predicted probabilities (6 output) as input features and fit a linear model on the validation data. Weights were optimized with genetic algorithms. I did the same for Jigsaw2019 data (7 output). The Ruddit data have only a single label so no optimization is done on the validation data. </p> <p>The table below shows the performance of all my models. I only used roberta and deberta models as they gave the best performance. Final submission is a weighted rank average of 15 models. "-l" means I included the duplicates between Jigsaw18 and validation.</p> <table> <thead> <tr> <th>model</th> <th>data</th> <th>validation</th> <th>public LB</th> <th>private LB</th> </tr> </thead> <tbody> <tr> <td>roberta-base</td> <td>jigsaw18</td> <td>0.7023</td> <td>0.7815</td> <td>0.8052</td> </tr> <tr> <td>roberta-large</td> <td>jigsaw18</td> <td>0.7035</td> <td>0.7788</td> <td>0.8064</td> </tr> <tr> <td>deberta-base</td> <td>jigsaw18</td> <td>0.7040</td> <td>0.7598</td> <td>0.8030</td> </tr> <tr> <td>deberta-large</td> <td>jigsaw18</td> <td>0.7050</td> <td>0.7906</td> <td>0.8139</td> </tr> <tr> <td>roberta-base-l</td> <td>jigsaw18</td> <td>0.7028</td> <td>0.7690</td> <td>0.8070</td> </tr> <tr> <td>roberta-large-l</td> <td>jigsaw18</td> <td>0.7027</td> <td>0.7737</td> <td>0.8013</td> </tr> <tr> <td>deberta-base-l</td> <td>jigsaw18</td> <td>0.7030</td> <td>0.7474</td> <td>0.8013</td> </tr> <tr> <td>deberta-large-l</td> <td>jigsaw18</td> <td>0.7044</td> <td>0.7716</td> <td>0.8085</td> </tr> <tr> <td>roberta-base</td> <td>jigsaw19</td> <td>0.7008</td> <td>0.7617</td> <td>0.8020</td> </tr> <tr> <td>roberta-large</td> <td>jigsaw19</td> <td>0.6991</td> <td>0.7468</td> <td>0.7968</td> </tr> <tr> <td>deberta-base</td> <td>jigsaw19</td> <td>0.7026</td> <td>0.7403</td> <td>0.7958</td> </tr> <tr> <td>roberta-base</td> <td>ruddit</td> <td>0.6859</td> <td>0.8108</td> <td>0.7845</td> </tr> <tr> <td>roberta-large</td> <td>ruddit</td> <td>0.6865</td> <td>0.8132</td> <td>0.7955</td> </tr> <tr> <td>deberta-base</td> <td>ruddit</td> <td>0.6880</td> <td>0.7903</td> <td>0.7880</td> </tr> <tr> <td>deberta-large</td> <td>ruddit</td> <td>0.6942</td> <td>0.8296</td> <td>0.7989</td> </tr> <tr> <td>ensemble</td> <td>jigsaw18</td> <td></td> <td>0.7763</td> <td>0.8103</td> </tr> <tr> <td>ensemble</td> <td>jigsaw19</td> <td></td> <td>0.7509</td> <td>0.8012</td> </tr> <tr> <td>ensemble</td> <td>ruddit</td> <td></td> <td>0.8235</td> <td>0.7983</td> </tr> <tr> <td>ensemble</td> <td>all15</td> <td></td> <td>0.7879</td> <td>0.8139</td> </tr> </tbody> </table> <p>At last I would like to thank Jigsaw and Kaggle for hosting this competition.</p> <p>Link to inference code: <a href="https://www.kaggle.com/wowfattie/notebook9298460840" target="_blank">https://www.kaggle.com/wowfattie/notebook9298460840</a><br> Link to training code: <a href="https://github.com/GuanshuoXu/Jigsaw-Rate-Severity-of-Toxic-Comments" target="_blank">https://github.com/GuanshuoXu/Jigsaw-Rate-Severity-of-Toxic-Comments</a></p>
Jigsaw Rate Severity of Toxic Comments	24th Place: Late Interaction + TPUs + More	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h2>Late Interaction Mechanism</h2> <p>The idea for this is loosely taken from the late interaction mechanism of <a href="https://arxiv.org/pdf/2004.12832.pdf" target="_blank">ColBERT</a> paper. <br> The task at hand is to rank 14,000 comments based on their level of toxicity. This is similar to search ranking on Google, where a model must quickly sort through millions of search results based on a query. I started reading papers on search ranking in the hope of finding something useful. This is where I came across Cross-encoders, Bi-encoders, and the Late interaction mechanism. </p> <p>I have made a post about <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/296349" target="_blank">Cross Encoders vs Bi Encoders</a> previously. Late interaction mechanism can be seen as combining the representation power of cross-encoders and the low latency of bi-encoders. </p> <p>We are given the training data in pair form of (less_toxic_comment, more_toxic_comment). </p> <p>A cross-encoder would be able to better model this pairwise interaction but it cannot be used in test time because there are total of 14k*14k/2 = 98M possible pair interactions between the comments.<br> On the other hand, a bi-encoder would only be able to look at each comment individually and would compress the comment into a vector.</p> <p>Here is where I use the late interaction mechanism to approximate all the pairwise interactions in a time-efficient way. This boosted my public LB from 0.844 to 0.872. Sadly it did not have such a big impact on private LB.</p> <h2>Other Tricks</h2> <ul> <li><strong>Psuedo Labeling</strong>: I used the 2018 competition dataset for pseudo-labeling which gave a big improvement in score</li> <li><strong>Muppet Roberta Large</strong>: I got better performance from Muppet Roberta Large backbone by Facebook, probably because it was already trained on sentiment and sentence similarity tasks.</li> <li><strong>Ensembling</strong>: Ensembling on a single backbone did not help because of soft labels of pseudo labels. On retrospect I should have ensembled more.</li> <li><strong>Loss Function:</strong> I used dice loss with alpha 0.25 from the paper <a href="https://arxiv.org/abs/1911.02855" target="_blank">Dice Loss for Data Unbalanced NLP Tasks</a> to train on the binary labels: severe_toxic, identity_hate, threat, toxic, obscene, and insult. I came across this paper while researching the Feedback Competition, but it also turned out to be useful for this competition. </li> <li><strong>TPUs</strong>: To speed up the training I used a batch size of 512 with mixed precision. With this training a Roberta large model on 200k+ comments took <5 mins / epoch. I generally use Tensorflow+TPU for NLP because 20 hours of free 8xTPUv3 >> 30 hours of free P100 GPU. </li> <li><strong>Hyperparameters</strong>: To make up for the large batch size I set beta_2 to 0.99 and used a cyclic learning rate with a max LR of 16e-5. </li> <li><strong>SWA</strong>: I also implemented a SWA callback in Keras to checkpoint the weights at the end of each cycle. I averaged the weights in the end (Model Soup).</li> <li><strong>Moving Average</strong>: Small lift with using EMA of 0.99</li> </ul> <p>I used some of the highest-scoring public notebooks for generating additional pseudo labels. I also trained a model on 2019 competition data to generate even more pseudo labels. This way, for each comment, I had 20+ labels. </p> <p>Round 0 public LB: 0.833 \| Round 1 public LB: 0.837 \| Round 2 public LB: 0.844 \| Round 3 public LB: 0.842 </p> <hr> <p>Hindsight: I really wish I had read this paper by Deepmind: <strong>Multiplicative Interactions and Where to Find Them</strong></p>
Jigsaw Rate Severity of Toxic Comments	27th Solution - What Worked and What Didn't for Me...	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>First of all congratulations, all for your hard work!</p> <p><strong>Overview</strong><br> It was a weird experience, CV LB gap was insane and really unstable so I stop looking at LB scores after my first couple submissions unless something extraordinary happens (like really high or low score), otherwise I just concentrated on my CV scores. Then the question appeared: "Which CV to trust?" So I started by choosing CV strategy first:</p> <p><strong>CV Strategy</strong><br> I tried several but at the end I decided to use <a href="https://www.kaggle.com/its7171" target="_blank">@its7171</a>'s CV strategy <a href="https://www.kaggle.com/its7171/jigsaw-cv-strategy" target="_blank">here</a>. It made sense to me and giving somewhat more stable results.</p> <p><strong>Data</strong></p> <ul> <li>Started with validation data only before getting into other datasets.</li> <li>Some of these others were pretty popular with other competitors. (Like old competition data or Ruddit etc.) then decided to rule some of them out to avoid possible leaks.</li> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/302060" target="_blank">Measuring Hate Speech</a> with some preprocessing. I created less and more toxic pairs using several permutations and keeping the closest scored ones in dataset and added samples from it to main pipeline without replacement.</li> </ul> <p><strong>Loss & Metric</strong></p> <ul> <li>Even though other loss functions like MSE etc. gave me decent CV scores (0.70+, 0.71+) at the end I decided to go with MarginRankingLoss.</li> <li>Metric wise I used what described in competition evaluation page…</li> </ul> <p><strong>Model</strong></p> <ul> <li><p>My final ensemble consist of transformer based models only. Which include,</p> <ul> <li>roberta variants mostly (Including distilbert),</li> <li>deberta v3 variants,</li> <li>bert variants (mostly toxic, hate ones)</li> <li>bart</li></ul></li> <li><p>Without external data my average CV was around <code>0.71</code> with external data included my CV didn't change much but a little bit less than the original scores but I decided to keep high scoring external data models to increase diversity at final ensemble.</p></li> <li><p>Models trained based on MarginRankingLoss, I didn't train the actual transformer blocks much (low learning rate for the backbones), instead I trained a final custom attention on top of their outputs and output layer.</p></li> <li><p>Fine-tuned these models with frequent validation checks. Usually converged around 3/4'th of first epoch.</p></li> </ul> <p>Final submission was weighted sum of the rankings from these models.</p> <p><strong>Things Might Have Worked</strong></p> <ul> <li>Models with sentence transformer embeddings. I created high dimensional matrices by concatenating different base transformer model embeddings. (This one was my second final selection and made barely into silver zone, didn't have time to include more models, it was only roberta and roberta-large)</li> <li>Combining different approaches at final ensemble (linear, transformer, embedding ones etc.)</li> </ul> <p><strong>Things Didn't Worked</strong> </p> <ul> <li>Linear Models with TF-IDF</li> <li>Models with word embeddings (Glove, fasttext etc.)</li> <li>Linear models with transformer embeddings combined together TF-IDF</li> <li>Pretraining masked language models on toxic datasets before finetuning on actual task.</li> <li>Some transformer approaches like Electra (I was expecting better CV score for discriminators)</li> </ul> <p>At the end some of the earlier ensembles made much better than my final selection but that's the luck part in competition :) As I said this competition was an interesting experience for me, learned bunch of stuff on the road…</p>
Jigsaw Rate Severity of Toxic Comments	41st solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks! <br> The number of data and the threshold of toxic or nontoxic may be different.</p> <p>Thanks!<br> In my case, weighted avg boost CV about 0.002~3 from simple avg.</p> <p>Thank you!</p> <p>Congradulations, all winners. <br> I learned a lot about NLP in this competition. thank you!</p> <p>Once the ranking is confirmed, I will be the competition master! 🙂</p> <h3>Overview</h3> <p><img src="https://user-images.githubusercontent.com/70784979/152980279-cb5b02a9-63f3-499a-a60b-2e6e9a668114.png" alt="image"></p> <h3>Data</h3> <p>I Created two types of extra data.</p> <ol> <li>binary label data. (toxic or nontoxic)</li> <li>more_toxic, less_toxic data</li> </ol> <h3>CV strategy</h3> <p><a href="https://www.kaggle.com/its7171/jigsaw-cv-strategy" target="_blank">Jigsaw CV strategy</a></p> <h3>Models</h3> <table> <thead> <tr> <th>model</th> <th>CV</th> <th>loss</th> </tr> </thead> <tbody> <tr> <td>microsoft/deberta-large</td> <td>0.7032</td> <td>BCE</td> </tr> <tr> <td>gpt2-medium</td> <td>0.6993</td> <td>BCE</td> </tr> <tr> <td>distilroberta-base</td> <td>0.7012</td> <td>MarginRankingLoss</td> </tr> <tr> <td>microsoft/deberta-base</td> <td>0.7043</td> <td>MarginRankingLoss</td> </tr> <tr> <td>xlnet-base-cased</td> <td>0.6997</td> <td>MarginRankingLoss</td> </tr> <tr> <td>robelta-large</td> <td>0.702</td> <td>MarginRankingLoss</td> </tr> <tr> <td>facebook/bart-base</td> <td>0.7027</td> <td>MarginRankingLoss</td> </tr> <tr> <td>Hate-speech-CNERG/dehatebert-mono-english</td> <td>0.6769</td> <td>MarginRankingLoss</td> </tr> <tr> <td>roberta-base</td> <td>0.7017</td> <td>MarginRankingLoss</td> </tr> <tr> <td>microsoft/deberta-large</td> <td>0.7014</td> <td>MarginRankingLoss</td> </tr> </tbody> </table> <p>The larger model had a better CV.</p> <p>max_len : 256</p> <p>optimizer : AdamW<br> lr : 1e-5 <br> weight_decay : 0.1<br> epochs : 1~3<br> scheduler : get_linear_schedule_with_warmup</p> <h3>Ensemble</h3> <p>I use optuna. As described below</p> <pre><code>cols = ['less_toxic_pred', 'more_toxic_pred'] oof = [] oof.append(pd.read_csv('../input/jigsaw4-model2/bi003_deberta-large/oof.csv',usecols=cols).values) oof.append(pd.read_csv('../input/jigsaw4-model1/exp008_roberta-base/oof.csv',usecols=cols).values) def calc_score(weight): pred = np.zeros([n_data,2]) for p, w in zip(oof,weight): pred += p*w score = (pred[:,0] < pred[:,1]).sum()/pred.shape[0] return score class Objective: def __init__(self, n_models): self.n_models = n_models def __call__(self, trial): weight = [trial.suggest_uniform('weight' + str(n), 0, 1) for n in range(self.n_models)] return calc_score(weight) max_iter = 1800 SEED = 29 objective = Objective(n_models) sampler = optuna.samplers.TPESampler(seed=SEED) study = optuna.create_study(sampler = sampler,direction='maximize') study.optimize(objective, n_trials = max_iter, n_jobs = -1) </code></pre>
Jigsaw Rate Severity of Toxic Comments	4th - This is Great! - Shared Solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks <a href="https://www.kaggle.com/titericz" target="_blank">@titericz</a>, </p> <p>It's great for me to see you write here!</p> <p>Hi <a href="https://www.kaggle.com/cdeotte" target="_blank">@cdeotte</a>, </p> <p>Thanks for your greeting!</p> <p><a href="https://www.kaggle.com/pdnartreb" target="_blank">@pdnartreb</a> and his thread <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/304441" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/304441</a> comments about <a href="https://www.kaggle.com/sorenj" target="_blank">@sorenj</a> and his notebooks <a href="https://www.kaggle.com/sorenj/perspectiveapi-tuning" target="_blank">https://www.kaggle.com/sorenj/perspectiveapi-tuning</a> and <a href="https://www.kaggle.com/sorenj/scoring-ruddit-comments" target="_blank">https://www.kaggle.com/sorenj/scoring-ruddit-comments</a></p> <p>In the first one you can see this piece of code,</p> <p><code>validation_annotations = validation.set_index('more_toxic').join( comments[['comment_id', 'text']].set_index('text'), how='inner').set_index('less_toxic').join( comments[['comment_id', 'text']].set_index('text'), lsuffix='_more', rsuffix='_less', how='inner').reset_index(drop=True)</code></p> <p>There you can follow the answer!</p> <p>Hi <a href="https://www.kaggle.com/virilo" target="_blank">@virilo</a>, thanks for your congratulations, I have just shared the complete solution if it answers your question!</p> <p>I vaguely referred to "cv" just to indicate the term validation in a generic way. Actually only "tito cv" is really cross validation in the solution. When I refer to "host cv" it is not technically a cross validation but only a selection of the host validation data. </p> <p>You can use this link <a href="https://scikit-learn.org/stable/modules/classes.html#module-sklearn.model_selection" target="_blank">https://scikit-learn.org/stable/modules/classes.html#module-sklearn.model_selection</a> to read more about other validation sets beyond this competition.</p> <p>Your thread finished opening my eyes!</p> <p>Thanks to you <a href="https://www.kaggle.com/pdnartreb" target="_blank">@pdnartreb</a> </p> <p>This is Great!</p> <p>I just got the results and I'm super happy.</p> <p>Thanks to everyone who participated in the competition, to the host who proposed the challenge and to kaggle for making it easy for me to use their GPU resources.</p> <p>I won't have much time today until the end of the day but I'll quickly share my solution here, which was a final combo of roberta base (tito cv 0.6983) + roberta large (tito cv 0.7014) with detoxify (host cv 0.7036).</p> <p>Thanks to everyone once again!</p> <p>UPDATED EDIT: </p> <h5>Shared Code</h5> <p>I just shared the solution in the following training and inference notebooks,</p> <p>Training:</p> <ul> <li><a href="https://www.kaggle.com/coreacasa/jigsaw4-luke-base-training-tito-cv-strategy/notebook" target="_blank">https://www.kaggle.com/coreacasa/jigsaw4-luke-base-training-tito-cv-strategy/notebook</a></li> <li><a href="https://www.kaggle.com/coreacasa/jigsaw4-luke-large-training-tito-cv-strategy/notebook" target="_blank">https://www.kaggle.com/coreacasa/jigsaw4-luke-large-training-tito-cv-strategy/notebook</a></li> </ul> <p>Inference:</p> <ul> <li><a href="https://www.kaggle.com/coreacasa/jigsaw4-merging-luke-roberta-with-detoxify/notebook?scriptVersionId=87076833" target="_blank">https://www.kaggle.com/coreacasa/jigsaw4-merging-luke-roberta-with-detoxify/notebook?scriptVersionId=87076833</a></li> </ul> <p>Acknowledgment and Credits to the authors in the comments section that I repeat here (I'm sorry if I didn't include someone): </p> <ul> <li><p><a href="https://www.kaggle.com/yasufuminakama" target="_blank">@yasufuminakama</a> and his notebooks <a href="https://www.kaggle.com/yasufuminakama/jigsaw4-luke-base-starter-train" target="_blank">https://www.kaggle.com/yasufuminakama/jigsaw4-luke-base-starter-train</a> and <a href="https://www.kaggle.com/yasufuminakama/jigsaw4-luke-base-starter-sub" target="_blank">https://www.kaggle.com/yasufuminakama/jigsaw4-luke-base-starter-sub</a></p></li> <li><p><a href="https://www.kaggle.com/its7171" target="_blank">@its7171</a> for their notebook <a href="https://www.kaggle.com/its7171/jigsaw-cv-strategy" target="_blank">https://www.kaggle.com/its7171/jigsaw-cv-strategy</a></p></li> <li><p><a href="https://www.kaggle.com/steubk" target="_blank">@steubk</a> and his dataset <a href="https://www.kaggle.com/steubk/detoxify-sourcemodels" target="_blank">https://www.kaggle.com/steubk/detoxify-sourcemodels</a></p></li> <li><p><a href="https://www.kaggle.com/pdnartreb" target="_blank">@pdnartreb</a> and his thread <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/304441" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/304441</a> </p></li> <li><p>(from previous discussion) <a href="https://www.kaggle.com/sorenj" target="_blank">@sorenj</a> and his notebooks <a href="https://www.kaggle.com/sorenj/perspectiveapi-tuning" target="_blank">https://www.kaggle.com/sorenj/perspectiveapi-tuning</a> and <a href="https://www.kaggle.com/sorenj/scoring-ruddit-comments" target="_blank">https://www.kaggle.com/sorenj/scoring-ruddit-comments</a></p></li> </ul> <p>UPDATED EDIT2: </p> <h5>Overview</h5> <p>Jigsaw Rate has continued the motivation of the 2018-2019-2020 challenges. Beyond the new metric (task) that evaluated the final performance of the solutions. I would add several peculiarities in this new challenge with respect to the previous ones, one of them (1) it is that the host did not provide a specific training data set, another (2) is that the host did provide a validation data set and finally (3) that we had a scarce reference of the public board scores representing only 5% of the final result.</p> <p>Once we advanced in the development of the competition and derived from (1) and (2) above, I thought that as important as the diversity of models, perhaps this time the diversity of training datasets could play an even stronger role. Likewise, derived from (2) and (3) and from the lack of gratitude that any submission to the public scoreboard was, I decided in the last week to give all priority to validation scores regardless of whether the public board confirmed improvements in performance or not.</p> <h6>Diversity of Datasets:</h6> <p>-From the previous Jigsaw competitions I worked with,<br> (2018 Jigasw1) - Toxic Comment Classification Challenge<br> (2019 Jigasw2) - Jigsaw Unintended Bias in Toxicity Classification<br> (2020 Jigsaw3) - Multilingual Toxic Comment Classification</p> <p>-From this competition,<br> (2021 Jigasw4) - 'validation_data.csv'.</p> <p>Although they are not part of the solution I also tried, without success, with external data. I'm referring to ruddit and hate_speech data, thanks <a href="https://www.kaggle.com/andre112" target="_blank">@andre112</a> and thanks <a href="https://www.kaggle.com/rajkumarl" target="_blank">@rajkumarl</a>. Specifically, I extracted samples of paired texts for each one of them given the enormous final extension that the combination of all the crosses supposed. Perhaps I made a mistake in this and I should have sampled the texts to work with the resulting paired texts in a number appropriate to my available resources.</p> <h6>Validation Strategies:</h6> <p>As I indicated in a brief reference at the beginning of this discussion, the final solution uses 2 validations,</p> <p>1) tito-cv (original author's nickname) applied to validation_data.csv of this competition. This strategy extends a 5-fold cross-validation by non-overlapping Groups of 4143 leak-free (least toxic-most toxic) text pairs (14251 different texts).</p> <p>2) host-cv is not cross-validation (sorry if I am confusing the vagueness of 'cv') but simply an extraction of text pairs (least toxic-most toxic) from the validation_data.csv dataset (same as above) that match with the dataset comments_to_score.csv, in total 8334 pairs found.</p> <h6>Models and Trainings:</h6> <p>Taking into account the datasets that are ingested, we can divide this part into two sections,</p> <p>1) Two Luke-Roberta models were fine-tuned on the validation_data.csv of this competition, please visit <a href="https://github.com/studio-ousia/luke" target="_blank">https://github.com/studio-ousia/luke</a>.</p> <p>I forked Y.Nakama's notebook Luke-base (Roberta-Base) by replacing the original validation pipeline (groupkfold over worker_id) with the aforementioned tito-cv. Saving the best of 15 epochs for each fold, I kept the rest of the parameters untouched (among others… max_len= 64, batch_size= 64, AdamW, lr's= 1e-5, 1/2 cycle cosine) as well as the loss function Margin-Ranking-Loss with margin= 0.5.</p> <p>I trained a second model this time changing the base model to Roberta-Large, only 5 epochs keeping everything else the same. On these 2 models, a result of linear optimized blending was tito-cv 0.705 (I share this new notebook <a href="https://www.kaggle.com/coreacasa/jigsaw4-luke-base-large-tito-cv-merging-optimize" target="_blank">here</a>) although I did not use this blend in the end.</p> <p>2) About the training data of the Jigsaw1 (original), Jigsaw2 (unbiased) and Jigsaw3 (multilingual) competitions I used (did not train) already trained models to predict toxic comments about those 3 competitions, please visit <a href="https://github.com/unitaryai/detoxify" target="_blank">https://github.com/unitaryai/detoxify</a>.</p> <table> <thead> <tr> <th>Model name</th> <th>Transformer type</th> <th>Data from</th> </tr> </thead> <tbody> <tr> <td>original</td> <td>bert-base-uncased</td> <td>Toxic Comment Classification Challenge</td> </tr> <tr> <td>unbiased</td> <td>roberta-base</td> <td>Unintended Bias in Toxicity Classification</td> </tr> <tr> <td>multilingual</td> <td>xlm-roberta-base</td> <td>Multilingual Toxic Comment Classification</td> </tr> </tbody> </table> <p>With a high performance for single models (not ensembles) in the leaderboards of each challenge and evaluating their individual performance on host validation, I considered it a great match to include these predictions in the final solution.</p> <p>As for the labels on which these models return a probability value, we would have,</p> <table> <thead> <tr> <th>Model name</th> <th>Labels</th> </tr> </thead> <tbody> <tr> <td>original</td> <td>{'toxicity', 'severe_toxicity', 'obscene', 'threat', 'insult', 'identity_attack'}</td> </tr> <tr> <td>unbiased</td> <td>{'toxicity', 'severe_toxicity', 'obscene', 'identity_attack', 'insult', 'threat', 'sexual_explicit'}</td> </tr> <tr> <td>multilingual</td> <td>{'toxicity', 'severe_toxicity', 'obscene', 'identity_attack', 'insult', 'threat', 'sexual_explicit'}</td> </tr> </tbody> </table> <p>The next step was to find a method to translate these multivariable outputs into a single output that measures toxicity, beyond the 'toxicity' label itself, and that would work with the 'Average Agreement with Annotators' in this competition. </p> <p>In the first of <a href="https://www.kaggle.com/sorenj" target="_blank">@sorenj</a> 's notebooks, the author uses an interesting answer on the above by proposing a random search of weights for each label in order to find a optimum in the metric for with host-cv. That same approach but with some nuances was the one I finally followed.</p> <h6>Combining All Things:</h6> <p>From the previous approach, I tried several alternatives to merge the detoxify's labels. Rather than looking for a solution close to the linear optimum (increasing the number of iterations), I looked at the variability of the result of each iteration on host-cv. Since nothing guaranteed that the texts in the private belonged to the same family as the validation ones, I thought it might be a good idea to try a non-optimized submission but simply play with the results of an iteration.</p> <p>To later combine the 3 models I did the same. The idea was not to optimize twice on the same texts and on the same workers. I also combined the luke's like this and therefore I did not use the training oofs but directly applying them to the prediction.</p> <p>Luke's final ensemble with detoxify's was with ranges and I didn't use maths or randomness this time. Simply the kaggle-art to balance dataset diversity and model diversity (1/3 for the former and 2/3 for the latter). A final action was to randomize the cases of duplicate scores in the final submission if applicable.</p> <h6>Summary of Results:</h6> <p>I repeat the well-known validation scores and add the evaluation on the public and private leaderboard,</p> <table> <thead> <tr> <th>Model name</th> <th>Validation</th> <th>Public</th> <th>Private</th> </tr> </thead> <tbody> <tr> <td>luke-base</td> <td>0.6983</td> <td>0.79993</td> <td>0.79167</td> </tr> <tr> <td>luke-large</td> <td>0.7014</td> <td>0.80287</td> <td>0.79898</td> </tr> <tr> <td>original</td> <td>0.69450</td> <td>0.74507</td> <td>0.79883</td> </tr> <tr> <td>unbiased</td> <td>0.69750</td> <td>0.73364</td> <td>0.79768</td> </tr> <tr> <td>multilingual</td> <td>0.69138</td> <td>0.73919</td> <td>0.79579</td> </tr> </tbody> </table> <p>Next step,</p> <table> <thead> <tr> <th>Ensemble</th> <th>Validation</th> <th>Public</th> <th>Private</th> <th>Weights (norm)</th> </tr> </thead> <tbody> <tr> <td>luke's</td> <td>0.75150(leaked)</td> <td>0.80439</td> <td>0.79698</td> <td>[0.53215757 0.46784243]</td> </tr> <tr> <td>detoxify's</td> <td>0.70362</td> <td>0.74605</td> <td>0.80791</td> <td>[0.38403517 0.43512979 0.18083503]</td> </tr> </tbody> </table> <p>Last step,</p> <table> <thead> <tr> <th>Submission</th> <th>Public</th> <th>Private</th> </tr> </thead> <tbody> <tr> <td>final ensemble</td> <td>0.77587</td> <td>0.81282</td> </tr> </tbody> </table> <h6>Checking Reproducibility and Randomness:</h6> <p>In case some strange phenomenon had been introduced in the solution, I checked <a href="https://www.kaggle.com/coreacasa/jigsaw4-checkup-reproducibility-4th-place-solution?scriptVersionId=87635669" target="_blank">here</a> its reproducibility.</p> <p>To check how important randomness (seed) is in the solution, I ran this <a href="Jigsaw4-Checkup-Randomness-4th-Place-Solution" target="_blank">this notebook</a> and tried various seed changes (or intermediate additions) keeping the results a gold standard.</p> <p>(In any case… yes, this time luck was on my side!)</p> <p>That's all folks!</p>
Jigsaw Rate Severity of Toxic Comments	52nd - Silver Medal - RAPIDS Forest Inference Library!	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks Patrick</p> <h1>Trust Your CV - 1800th Public to 52nd Private!</h1> <p>Thanks for a great competition. Public leaderboard was very misleading since it only contained 5% of test data. For my final two submissions, I chose two submissions that were ranked 1500th and 1800th on public LB. They both finished at 52nd private. Each had my best leak-free CV score of 0.707, public LB 0.802, private LB 0.804!</p> <h1>RAPIDS FIL infers 100 million rows per second!</h1> <p>In this competition, i chose to use RAPIDS Forest Inference Library because it can infer up to 100 million rows of data per second! See blog <a href="https://medium.com/rapids-ai/rapids-forest-inference-library-prediction-at-100-million-rows-per-second-19558890bc35" target="_blank">here</a></p> <h1>Feature Engineering - roBERTa XGB</h1> <p>For each comment, we extract 768 features from <code>roBERTa-base</code> that was pretrained on old comp data with old comp targets. And we extract 1024 features from <code>roBERTa-large</code> that was pretrained on old comp data with old comp targets. Next we train XGB using these features and infer with RAPIDS FIL. We did not use Tfidf nor linear regression models.</p> <h1>Infer Every Possible Pair - RAPIDS FIL</h1> <p>Since inference is so fast, instead of predicting 14,000 toxic scores, we can predict 196,000,000 million combinations of every pair of toxic test comments. During inference we build a RAPIDS cuDF dataframe with <code>196,000,000</code> million rows and <code>3584</code> columns which represents every possible pair <code>14,000 * 14,000 = 196,000,000</code>. Then we use RAPIDS FIL to predict every row and get 196,000,000 probabilities.</p> <h1>Convert Pair Probabilities into Toxic Scores</h1> <p>Based on the 196,000,000 comparisons of each of the 14,000 test comments with every other comment, we compute toxic scores for each of the 14,000 test comments and merge it unto <code>submission.csv</code>.</p> <h1>Training Code and Inference Code</h1> <p>I published train code <a href="https://www.kaggle.com/cdeotte/rapids-xgb-fil-silver-medal-train-0-804" target="_blank">here</a>. I published infer code <a href="https://www.kaggle.com/cdeotte/rapids-xgb-fil-silver-medal-infer-0-804" target="_blank">here</a></p> <p>Thanks vad13irt</p> <p>Thanks. Train code is <a href="https://www.kaggle.com/cdeotte/rapids-xgb-fil-silver-medal-train-0-804" target="_blank">here</a> and infer code is <a href="https://www.kaggle.com/cdeotte/rapids-xgb-fil-silver-medal-infer-0-804" target="_blank">here</a>. Training is only 11 folds on 30,000 rows of validation data. Inference is 196,000,000 rows on all possible pairs of test comments.</p> <p>Thanks <a href="https://www.kaggle.com/datafan07" target="_blank">@datafan07</a> . Congrats on your great finish. RAPIDS FIL is incredible. I've been using it at work and wanted to try it in a competition.</p> <p>Whenever you have a tree model (XGB, LGBM, Sklearn Random Forest, whatever), you can load it with RAPIDS FIL and infer. If you infer XGB as FIL on GPU it is like 10x faster than inferring XGB as XGB on GPU! So it isn't just about GPU, the FIL library evaluates the tree splits more efficiently than any of the tree libraries by themselves. </p> <p>Congrats on achieving silver medal.</p> <p>It will take between 1-7 days. Kaggle will finalize the private LB by removing teams that violated rules. Afterward, the comp will be finalized and you get your medal.</p> <p>Thanks. Are you asking about my journey in this competition? Or my life's data science journey?</p>
Jigsaw Rate Severity of Toxic Comments	5th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>I decided to focus on maximizing only validation score in the last month of the competition. Public LB did not seem quite useful because it had only 5% of data and there was not correlation between validation score and public LB score. <br> I used 2 datasets for training models: Ruddit dataset, Toxic comments dataset (<a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge)" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge)</a>. I tried other datasets, but they did not improve my validation score at all.<br> I used “validation_data.csv” for validation only.</p> <h2>Ruddit dataset / models</h2> <p>Ruddit dataset includes tuples (A, B, C, D) where “A” is the least toxic and “D” is the most toxic. Therefore, I used those tuples to create pairs (X, Y) where X is less toxic than Y. <br> Then I trained 2 models with MarginRankingLoss:<br> 1) “roberta-large” with linear layer <br> 2) “unitary/unbiased-toxic-roberta” with linear layer<br> I used layerwise learning rate decay (from 1e-5 to 1e-4). Training was made by using 5-folds cross-validation. Each model was trained with 3 epochs, after each epoch I saved the model with the best accuracy.</p> <h2>Toxic comments dataset / models</h2> <p>I used that dataset to train a regression models. Data preparation: 1) remove items that have the same text in validation_data.csv 2) compute toxic target (y = toxic<em>0.32 + severe toxic</em>1.82 + obscene<em>0.16 + threat</em>1.5 + insult<em>0.64 + identity_hate</em>1.5) 3) create a train dataset with positive toxic target and negative toxic target.<br> After data preparation I trained 3 models:<br> 1) TfIdf + Ridge. I used cleaned text and 4 extra features: length, portion of punctuation in text, portion of chars with uppercase in text, count of swear words (these features slightly improve score on validation_data). I trained models on 8 folds with grid search of hyperparameters ( “ngram_range”, “alpha”, “fit_intercept”) and chose the model with the highest score on validation_data.<br> 2) Regression model based on “unitary/unbiased-toxic-roberta” with linear custom layer and MSELoss. I trained that model by using 5-folds cross-validation with 5 epochs. The model with the minimal loss on validation fold was saved. Models were trained with layerwise learning rate. <br> 3) Regression model based on “unitary/unbiased-toxic-roberta” with attention custom layer and MSELoss. I trained that model by using 5-folds cross-validation with 5 epochs. The model with the minimal loss on validation fold was saved. Models were trained with layerwise learning rate.</p> <h2>Models ensemble</h2> <p>After training the models I used optuna to weigh models in ensemble. Steps:<br> 1) normalizing logits of models using min, max scaling 2) weighing models using weights 0 or 1. <br> After some experimentation I got maximum validation score - 0.709 with public LB score 0.819. <br> Final ensemble – sum of the models:</p> <table> <thead> <tr> <th><strong>№</strong></th> <th><strong>Model</strong></th> <th><strong>Dataset</strong></th> <th><strong>Count of models</strong></th> </tr> </thead> <tbody> <tr> <td>1</td> <td>roberta-large with MarginRankingLoss</td> <td>Ruddit</td> <td>2</td> </tr> <tr> <td>2</td> <td>unitary/unbiased-toxic-roberta with MarginRankingLoss</td> <td>Ruddit</td> <td>3</td> </tr> <tr> <td>3</td> <td>Tfidf + RIDGE</td> <td>Toxic comments</td> <td>3</td> </tr> <tr> <td>4</td> <td>unitary/unbiased-toxic-roberta with custom linear layer with MSELoss</td> <td>Toxic comments</td> <td>3</td> </tr> <tr> <td>5</td> <td>unitary/unbiased-toxic-roberta with custom attention layer with MSELoss</td> <td>Toxic comments</td> <td>2</td> </tr> </tbody> </table> <h2>Notebook</h2> <p>Link to my submission - <a href="https://www.kaggle.com/alexander1980/best-cv-ens?scriptVersionId=87213915" target="_blank">https://www.kaggle.com/alexander1980/best-cv-ens?scriptVersionId=87213915</a><br> P.S.: Sorry for duplications in my code.</p> <p><a href="https://www.kaggle.com/asarvazyan" target="_blank">@asarvazyan</a> , thank you!</p>
Jigsaw Rate Severity of Toxic Comments	62nd Place solution and what I learnt and tried	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>First of all I am pretty new in Kaggle Just started with Ventilator being my first real comp . Also learning along the way with great people like Chris, Abhishek and lot of great folks that share knowledge and great notebooks .</p> <p>My approach here might now be the most unique but still want to share my journey ..</p> <ul> <li><p>Started with roberta models trying to tune and getting the most juice of of those .. Got some good tips from <a href="https://www.kaggle.com/readoc" target="_blank">@readoc</a> </p></li> <li><p>Got a great suggestion for deberta models by <a href="https://www.kaggle.com/nbroad" target="_blank">@nbroad</a> and started to explore that . Though should have paid more attention to these as had great val scores and maybe would have finished higher . </p></li> <li><p>Started to train XLNET , distilbert and ensemble .Got some good lb</p></li> <li><p>Saw LB dropping with ensembles so selfishly tried to ensemble with linear models .</p></li> <li><p>AS the lb was very confusing so started to think to trust cv . developed a ensemble cv strategy by union folds for CV . Trusted it and ensembled based on it bu slowly adding models .</p></li> <li><p>Realized that Bert models really were good to cv so started adding that .</p></li> <li><p>Tried to train on ruddit , toxic data from old comps with some datasets people shared with MSE loss . Trained distil, xlnet , deberta and roberta models on this and added to ensemble .Some models with leaky relu worked good for distil and in some tinkered with various archs. Ruddit data scored less in public LB but ensembled well . Added one linear model using FastText with ridge from a public NB .</p> <ul> <li>For 1-2 of my ext data model did also tried to augment data a bit which helped cv for my distilbert models <br> distilbert old_toxic distil-tx-old mse 0.013031223 0.633618972 without Aug<br> distilbert_aug_relu old_toxic distil-tx-old mse 0.013226555 0.648996944 with Aug<br> distilAug_LRelu old_toxic distil-tx-old mse 0.013024052 0.6599907 WIth Aug<br> All ideas for NLP augs came from this notebook <a href="https://www.kaggle.com/shonenkov/nlp-albumentations" target="_blank">https://www.kaggle.com/shonenkov/nlp-albumentations</a> </li></ul></li> <li><p>Finally last day though to give deberta v3 another shot and added to ensemble . gave good cv but forgot to select it :D .</p></li> <li><p>Also tried weighted ensemble with optuna for tuning weights like people already shared . This would have made me land higher but somewhat didn't trust it so went with a simple average ensemble . Had used the validation data which I split up and scored and then put back in to get the score . Think many would have used like this only for weighted ensembles .</p></li> </ul> <p>Final ensembles and one selected in bold </p> <table> <thead> <tr> <th>ENS</th> <th>TYPE</th> <th>LB</th> <th>CV</th> <th>PRIVATE</th> </tr> </thead> <tbody> <tr> <td><strong>Ens with relu with toxic and ruddit data</strong></td> <td>Avg</td> <td>87.6</td> <td>70.46728972</td> <td>0.80434</td> </tr> <tr> <td>Ens with deberta</td> <td>Avg</td> <td>87.1</td> <td>70.99</td> <td>0.80634</td> </tr> <tr> <td>weighted ens without deb</td> <td>Weighted</td> <td>86.785</td> <td>72.1528</td> <td>0.80615</td> </tr> </tbody> </table> <p><strong>Models tried</strong><br> deberta v3<br> distil<br> roberta<br> xlnet</p> <p><strong>datasets tried</strong> <br> ruddit<br> clean validation data<br> old comp data .</p> <p>Finally went with deberta distil and roberta . .and distil and roberta for final ensemble </p> <p>Anyways was good to learn to structure experiment better at the end and have a good trusted cv code. Which I realized pretty late and maybe had missed lot of great models I had tried since the start of the comp . </p> <p>But at the end there was always new things to learn and learn to improve .. Thanks to all </p> <p>Thanks to you .. was a great suggestion and most top rakers think tried it. Though the low public lb score made me lose confidence in it initially which was a bad decision. Must have had more confidence in it and it's good cv 😊</p>
Jigsaw Rate Severity of Toxic Comments	6th Place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h2>Overview</h2> <p>Congradulations, all winners. <br> But really lucky for us, We got a gold medal. None of our members could have imagined this result. As pointed out <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/306366" target="_blank">in 7th place solution</a>, we had a lot of models that were trained mainly on Jigsaw 1st directly, and it is possible that the performance was better for private tests. This was simply luck.</p> <h2>Data</h2> <ul> <li><a href="https://www.kaggle.com/julian3833/jigsaw-toxic-comment-classification-challenge" target="_blank">jigsaw-toxic-comment-classification-challenge(jigsaw_1)</a></li> <li><a href="https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset" target="_blank">ruddit jigsaw dataset</a></li> <li><a href="https://www.kaggle.com/nkitgupta/jigsaw-regression-based-data" target="_blank">jigsaw_regression_data</a></li> </ul> <p>When we used jigsaw_1 to train a regression models, we calculated target values by following formula.</p> <p><code>(y = toxic0.32 + severe toxic1.82 + obscene0.16 + threat1.5 + insult0.64 + identity_hate1.5)</code></p> <h2>CV strategy</h2> <p>We used validation_data.csv to evaluate cv score. (In order to evaluate more accurately,) We removed all sentences from the training data that overlapped with validation_data.</p> <h2>Preprocessing / Engineering</h2> <h3><i>mogmog part</i></h3> <p>The training data were jigsaw1(applied under sampling) and Ruddit. We apply multiple patterns of text preprocessing. After preprocessing, TF-IDF (and SVD for lgb) , gensim_embedding, Basic feature(Word count, Character count), FastText embedding were applied on each of them to generate features.</p> <h3><i>kma### part</i></h3> <p>The training data were jigsaw_1, ruddit, jigsaw_regression_data. we applied text cleaning on jigsaw_1 and ruddit, and none on jigsaw_regression_data.<br> we applied TF-IDF for each of them, and trained Ridge reg. (alpha=0.5, 1, 2).<br> In inference, alpha=0.5, 1, and 2 were used as an ensemble.</p> <h3><i>kfsky part</i></h3> <p>The training data was jigsaw1 data (applied under sampling, text cleaning) for training data and generate features.</p> <ul> <li>TF-IDF(and SVD)</li> <li>CountVectorizer(and SVD)</li> <li>Word count</li> <li>Character count</li> <li>FastText（100dim）</li> </ul> <p>Train LightGBM with these features.</p> <p>And we used TF-IDF and trained Ridge model.</p> <h2>Training Model</h2> <p>we trained following models.</p> <table> <thead> <tr> <th>Model</th> <th>Dataset</th> <th>Text Cleaning</th> <th>Feature</th> <th>CV</th> </tr> </thead> <tbody> <tr> <td>LightGBM_kfksy_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>TF-IDF, CountVectriser, FastText, Word count, Character count</td> <td>0.6843</td> </tr> <tr> <td>Ridge_kfksy_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>TF-IDF</td> <td>0.6639</td> </tr> <tr> <td>Ridge_kma###_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>TF-IDF</td> <td>0.6823</td> </tr> <tr> <td>Ridge_kma###_2</td> <td>ruddit</td> <td>Cleaned</td> <td>TF-IDF</td> <td>0.6312</td> </tr> <tr> <td>Ridge_kma###_3</td> <td>jigsaw_regression_data</td> <td>Non</td> <td>TF-IDF</td> <td>0.6713</td> </tr> <tr> <td>Ridge_mogmog_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>TF-IDF</td> <td>0.6810</td> </tr> <tr> <td>Ridge_mogmog_2</td> <td>ruddit</td> <td>Cleaned</td> <td>TF-IDF</td> <td>0.6283</td> </tr> <tr> <td>lightGBM_mogmog_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>gensim_Embedding & FastTextEmbedding & Basic_Feat</td> <td>0.6792</td> </tr> <tr> <td>lightGBM_mogmog_2</td> <td>ruddit</td> <td>Cleaned</td> <td>gensim_Embedding & FastTextEmbedding & Basic_Feat</td> <td>0.6572</td> </tr> <tr> <td>lightGBM_mogmog_3</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>TF-IDF & SVD</td> <td>0.6711</td> </tr> <tr> <td>lightGBM_mogmog_4</td> <td>ruddit</td> <td>Cleaned</td> <td>TF-IDF & SVD</td> <td>0.6130</td> </tr> <tr> <td>RoBERTa_base_mogmog_1</td> <td>jigsaw_1</td> <td>Cleaned</td> <td>Only text cleaning</td> <td>0.6879</td> </tr> </tbody> </table> <h2>Weight Ensembling</h2> <p>We ensembled these models. Before ensemble, we applied MinMaxScaler because the scale of the prediction values were different among the models. The weights of each model are determined so that the CV score is maximized. (we used <code>scipy.optimize.minimize(method='powell')</code>) . After ensembling, cv score was 0.7048.</p> <h2>Closing</h2> <p>Thank you <a href="https://www.kaggle.com/ryo1993" target="_blank">mogmog</a>, <a href="https://www.kaggle.com/kmakmar" target="_blank">kma###</a>, <a href="https://www.kaggle.com/mkt0309" target="_blank">mkt0309</a> for teaming up with us.</p> <h2>Clean Inference Code</h2> <p><a href="https://www.kaggle.com/ryo1993/inference-krkm-final-001?scriptVersionId=88284870" target="_blank">https://www.kaggle.com/ryo1993/inference-krkm-final-001?scriptVersionId=88284870</a></p>
Jigsaw Rate Severity of Toxic Comments	76 place and Lesson Learned	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Private Score 0.80433<br> Public Score 0.72950<br> <a href="https://www.kaggle.com/crischir/ktrain-regression-transformer?scriptVersionId=87187729" target="_blank">https://www.kaggle.com/crischir/ktrain-regression-transformer?scriptVersionId=87187729</a><br> First of all: trust your CV and your work <br> Second do not copy notebooks like "very simple code with score 0.885" - it was a lot of knowledge for tha script but it was not mine. From a public score of 0.87950 the final score was 0.74882<br> There are some questions regarding this big shake-up of the public leaderboard.<br> 1.Transformers vs Linear.</p> <ul> <li>Linear regression has a stellar performance on public LB. 0.9 using TF-IDF. <br> -Transformers were performing very poor in public LB. 0.72950 in public 0.8 in private LB<br> -For similar transformers models, the public LB was inconsistent: public score betweeen 0.72950 and 0.78230 with similar results in private LB: 8.01 even if the CV was the same (0.67-0.69)<br> 2.Scoring <br> Basically if someone has the original corpus from <a href="https://www.perspectiveapi.com" target="_blank">https://www.perspectiveapi.com</a> (<a href="https://github.com/conversationai/perspectiveapi/blob/main/example_data/perspective_wikipedia_2k_score_sample_20180829.csv" target="_blank">https://github.com/conversationai/perspectiveapi/blob/main/example_data/perspective_wikipedia_2k_score_sample_20180829.csv</a>)<br> my script does nothing :)<br> If someone add the probabilities from the initial jigsaw corpus the results are better.<br> My question remains:<br> Why the public LB was so different, if we compare with CV and Private LB scores. The private LB is corelated with CV. </li> </ul>
Jigsaw Rate Severity of Toxic Comments	7th Place Solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Our team didn't trust public LB, and worked towards improving the score on the un-processed validation_df. In the end, we improved the score to 0.7211 without leaks, which was the 7th best score in private LB. We thought this was a great result because we kept running Trust CV, but we later realized that we had ensembled models that had won the lottery.</p> <p>As other teams have pointed out, the private LB dataset's test comments overlapped with jigsaw's first competition, the Toxic Comment Classification Challenge. Therefore, it was possible to create a model within the gold zone by training this dataset directly.</p> <p>Our team placed in the gold zone, partly because we did a Trust CV, but essentially because we had models that won the lottery and we gave them enough weight in the weighted average. </p> <p>Here's a brief overview of what we've been working on since.</p> <h1>Model × Dataset</h1> <p>The following table shows the models used for the final submission, the dataset trained, and the scores. The "pseudo label" refers to the model trained with validation_df in toxic-xlm-roberta and pseudo-labeled on the dataset. This pseudo-labeling greatly improved the CV, while private LB did not.</p> <p>In addition, we strongly emphasized the importance of not leaking when training. Specifically, when training validation_data, union-find GroupKFold was used, and when training other datasets, the text in validation_data was removed.</p> <table> <thead> <tr> <th>name</th> <th>model</th> <th>dataset</th> <th>CV</th> <th>PublicLB</th> <th>PrivateLB</th> </tr> </thead> <tbody> <tr> <td>mst029</td> <td>toxic-xlm-roberta</td> <td>Ruddit (Pseudo Label)</td> <td>0.7126</td> <td>0.77391</td> <td>0.79796</td> </tr> <tr> <td>mst030</td> <td>toxic-xlm-roberta</td> <td>jigsaw 1st (Pseudo Label)</td> <td>0.7224</td> <td>0.77587</td> <td>0.79676</td> </tr> <tr> <td>msttweet</td> <td>toxic-xlm-roberta</td> <td>Toxic Tweet (Pseudo Label)</td> <td>0.7155</td> <td>0.77587</td> <td>0.79641</td> </tr> <tr> <td>colum014</td> <td>roberta-base</td> <td>validation_df</td> <td>0.70307</td> <td>0.79590</td> <td>0.79410</td> </tr> <tr> <td>colum015</td> <td>roberta-large</td> <td>validation_df</td> <td>0.70307</td> <td>0.80189</td> <td>0.79744</td> </tr> <tr> <td>colum016</td> <td>deberta-v3-base</td> <td>validation_df</td> <td>0.70646</td> <td>0.78730</td> <td>0.79594</td> </tr> <tr> <td>colum018</td> <td>unbiased-toxic-roberta</td> <td>validation_df</td> <td>0.70426</td> <td>0.80341</td> <td>0.79416</td> </tr> <tr> <td>colum019</td> <td>toxic-xlm-roberta</td> <td>validation_df</td> <td>0.70898</td> <td>0.76749</td> <td>0.79644</td> </tr> <tr> <td>colum020</td> <td>toxic-bert</td> <td>validation_df</td> <td>0.70456</td> <td>0.77533</td> <td>0.79422</td> </tr> <tr> <td>calpis001</td> <td>roberta-large</td> <td>jigsaw 1st</td> <td>0.7031</td> <td>0.79111</td> <td>0.81814</td> </tr> <tr> <td>calpis011</td> <td>toxic-xlm-roberta</td> <td>jigsaw 1st (Pseudo Label)</td> <td>0.7205</td> <td>0.77152</td> <td>0.79532</td> </tr> <tr> <td>calpis012</td> <td>toxic-xlm-roberta</td> <td>Toxic Tweet (Pseudo Label)</td> <td>0.7141</td> <td>0.76967</td> <td>0.79743</td> </tr> <tr> <td>naoism1004</td> <td>roberta-large</td> <td>jigsaw 1st</td> <td>0.7048</td> <td>0.80581</td> <td>0.81520</td> </tr> <tr> <td>naoism4001</td> <td>toxic-xlm-roberta</td> <td>jigsaw 1st (Pseudo Label)</td> <td>0.7220</td> <td>0.76325</td> <td>0.79824</td> </tr> <tr> <td>naoism4003</td> <td>toxic-xlm-roberta</td> <td>Civil Comment Dataset (Pseudo Label)</td> <td>0.7103</td> <td>0.76249</td> <td>0.79584</td> </tr> </tbody> </table> <h1>Ensemble</h1> <p>We used these models to perform the ensemble of weighted averages. The weights were determined using Optuna, which was implemented by <a href="https://www.kaggle.com/calpis10000" target="_blank">@calpis10000</a> .</p> <p>The implementation is as follows. We used the array of less and more predictions corresponding to validation_df to automatically determine the weights for each model that would score best. Also, there was a problem that the scales of the predictions for each model were different in this case. Therefore, we tried MinMaxScaler, StandardScaler, RankGauss, and so on. The best one was StandardScaler.</p> <pre><code>def val(trial): p1 = np.zeros(30108) p2 = np.zeros(30108) x0 = trial.suggest_float(pred_columns[0], 0, 1) x1 = trial.suggest_float(pred_columns[1], 0, 1) x2 = trial.suggest_float(pred_columns[2], 0, 1) x3 = trial.suggest_float(pred_columns[3], 0, 1) x4 = trial.suggest_float(pred_columns[4], 0, 1) x5 = trial.suggest_float(pred_columns[5], 0, 1) x6 = trial.suggest_float(pred_columns[6], 0, 1) x7 = trial.suggest_float(pred_columns[7], 0, 1) x8 = trial.suggest_float(pred_columns[8], 0, 1) x9 = trial.suggest_float(pred_columns[9], 0, 1) x10 = trial.suggest_float(pred_columns[10], 0, 1) x11 = trial.suggest_float(pred_columns[11], 0, 1) x12 = trial.suggest_float(pred_columns[12], 0, 1) x13 = trial.suggest_float(pred_columns[13], 0, 1) x14 = trial.suggest_float(pred_columns[14], 0, 1) x_list = [x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14] for i, x in enumerate(x_list): p1 += scaler_less[:, i] * x p2 += scaler_more[:, i] * x score = (p1 < p2).mean() return score optuna.logging.disable_default_handler() study = optuna.create_study( direction='maximize', sampler=optuna.samplers.TPESampler(seed=42) ) study.optimize(val, n_trials=1000) display(study.best_trial.number) display(study.best_trial.values) display(study.best_trial.params) </code></pre> <p>The weights are as follows. It was good to see that much greater weight was given to calpis001 and naoism1004. The results were such that highly correlated predictions didn't give as much weight to each other, which was consistent with our goal of finding a weighted average that would produce diversity.</p> <pre><code>{ 'mst029_toxic': 0.14599159334648742, 'mst030_toxic': 0.9400258577765264, 'msttweet_toxic': 0.2296755364508963, 'colum014_toxic': 0.46435656956718, 'colum015_toxic': 0.8156998001053446, 'colum016_toxic': 0.6003326557886719, 'colum018_toxic': 0.08731982956387017, 'colum019_toxic': 0.02050220970508744, 'colum020_toxic': 0.8002815649764881, 'calpis001_toxic': 0.8907374041352457, 'calpis011_toxic': 0.46794326158461463, 'calpis012_toxic': 0.0613928819716386, 'naoism1004_toxic': 0.9998908148400331, 'naoism4001_toxic': 0.9085892865154106, 'naoism4003_toxic': 0.0312185722129308 } </code></pre> <p>The above is our solution. Thank you <a href="https://www.kaggle.com/mst8823" target="_blank">@mst8823</a> , <a href="https://www.kaggle.com/calpis10000" target="_blank">@calpis10000</a> , and <a href="https://www.kaggle.com/naoism" target="_blank">@naoism</a> for teaming up with us!</p> <p>Inference code: <a href="https://www.kaggle.com/columbia2131/jigsaw-team-ensemble-006-fix/notebook" target="_blank">[Jigsaw] team-ensemble 006 fix</a></p> <p>Thank you for your comment!<br> But we didn't know if Optuna's tuning was good this time until we saw the results. <br> I wonder how much less trouble it would have caused if the public LB was reliable and the distributions of validation_df and test's dataset were close…</p> <p>Thank you for your comment!<br> Optuna is quite effective in this case of optimization for validation_df.<br> It may also be useful for determining the weight of the models and be good to have it as a handy tool👍</p>
Jigsaw Rate Severity of Toxic Comments	84th Place Approach (Trust the right CV)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>TLDR</p> <p>First The validation data was cleaned using <a href="https://www.kaggle.com/samarthagarwal23" target="_blank">@samarthagarwal23</a> 's notebook . <a href="" target="_blank">here</a></p> <p>Then the transformer models were trained for 5 fold on the clean data using ranking loss.</p> <table> <thead> <tr> <th>Model Name</th> <th>CV</th> <th>Public LB Score</th> </tr> </thead> <tbody> <tr> <td>roberta-base</td> <td>0.74</td> <td>0.811</td> </tr> <tr> <td>roberta-large</td> <td>0.76</td> <td>0.810</td> </tr> <tr> <td>deberta-base</td> <td>0.75</td> <td>0.806</td> </tr> <tr> <td>deberta-large</td> <td>0.79</td> <td>0.808</td> </tr> <tr> <td>albert-base</td> <td>0.73</td> <td>0.806</td> </tr> <tr> <td>sentence-transformer(mpnetv2)</td> <td>0.74</td> <td>-</td> </tr> <tr> <td>mpnet-base</td> <td>0.74</td> <td>-</td> </tr> </tbody> </table> <p>I didnt pay much heed to the public LB scores as I couldn't find any correlation. I did use linear models too but due to lack of any correlation (For example: TF-IDF 7-fold & & Fasttext ensemble on JTCC Data: 0.77 CV 0.86 LB . Hence no correlation).</p> <p>I also tried another method of training on the JTCC data but rather than combining the different category scores i trained models on each of the categories (toxic, threat, obscene etc.) using weighted BCELoss. On running a loop to obtain the best weights for combining the different category model predictions on the untouched validation_data I observed that almost all weightage was given to the "toxic" column so i stuck to training for the BCELoss for the "toxic" column only.</p> <p>I tried retraining the models pretrained on the JTCC data on the cleaned validation data in the ranking loss which gave me a small boost in CV.</p> <p>My final ensemble included only the base transformer models. which were ensembled with 7-fold TF-IDF and Fasttext models from <a href="https://www.kaggle.com/yamqwe/ridge-fasttext-improved-roberta" target="_blank">here</a>.</p> <p>Final Notebook Link: <a href="https://www.kaggle.com/aishikai/js-final-ensemble-0-813-lb-0-839-cv" target="_blank">here</a></p> <p>In this competition one major help to me was <a href="https://www.kaggle.com/samarthagarwal23" target="_blank">@samarthagarwal23</a> 's notebook which gave me a lot of insight into the data and finally helped me choose my sub.</p>
Jigsaw Rate Severity of Toxic Comments	99th place solution - silver	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks for a competition!</p> <p>After joining the competition and checking public solutions i made a few decisions:</p> <ol> <li>I will not use public lb for validation, just because it contains only 5% of data.</li> <li>Ill try linear models on validation data and find out, that they performed bad, so i focused on transformers.</li> </ol> <p><strong>Datasets:</strong></p> <ol> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data</a></li> <li><a href="https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data" target="_blank">https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/data</a></li> <li><a href="https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset" target="_blank">https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset</a></li> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/data" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating/data</a></li> </ol> <p>From "1", "2", "3" datasets i made a big one in some different ways:</p> <ol> <li>Thresholding datasets "2" features, converting it to format of dataset "1". Then convert "1"+"2" to toxicity_score by calculating mean score, concatenate it with "3".</li> <li>Calculate mean scores of dataset "1", then concatenate with "2" and "3".</li> </ol> <p>Implementation of both methods you can see in code below (i called them "sparse" and "dense" datasets).</p> <p>Dataset "4" (validation_data.csv) i used for validation. I didnt use cross validation, because train dataset consists of 2 195 487 samples and training one model on google colab takes from 20 to 50 hours.</p> <p>I select 3 models, trained with different datasets, loss_fuctions and sequence lengths. Two of them based on GroNLP/hateBERT, the last one cardiffnlp/twitter-roberta-base-hate. <br> It give me 0.79784 private score (0.6988 score on validation, was my first submission, 157 place and bronze zone).</p> <p>For the second submission i mixed models with other open solutions same way. But this two solutions are ensembles themselfs.</p> <ol> <li><a href="https://www.kaggle.com/yuzhoudiyishuai/robertabase5fold2-linear-256" target="_blank">https://www.kaggle.com/yuzhoudiyishuai/robertabase5fold2-linear-256</a></li> <li><a href="https://www.kaggle.com/saurabhbagchi/pytorch-w-b-jigsaw-starter" target="_blank">https://www.kaggle.com/saurabhbagchi/pytorch-w-b-jigsaw-starter</a></li> </ol> <p>So the weights: model_1, model_2, model_3, 1st ensemble, 2nd ensemble - each for 20% contributions in final result.</p> <p>Validation score of 5 models: 0.7190. It was my second submission, 0.80153 on private and 99th place.</p> <p>Full code of my solution with model weights and description you can find here:<br> <a href="https://github.com/vilka-lab/JigsawRate" target="_blank">https://github.com/vilka-lab/JigsawRate</a></p>
Jigsaw Rate Severity of Toxic Comments	A Detoxify solution (#3)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>congratulations to you too for your smart solution <a href="https://www.kaggle.com/coreacasa" target="_blank">@coreacasa</a> ! In any case Detoxify is a <a href="https://github.com/unitaryai/detoxify" target="_blank">github project by Unitaryai</a> that I have just assembled into a Kaggle dataset! 😊</p> <p>Thank you to Jigsaw and Kaggle for hosting a challenging and engaging competition and congratulations to the winners and everyone who enjoied the competition !</p> <p>I particularly enjoied this competition format because it forces to find a solution that generalizes on unseen data.</p> <h2>My Solution</h2> <p>For a solution that generalizes on unseen data I tried to use: </p> <ul> <li>datasets from previous Jigsaw competitions to extract features</li> <li>VAL data + RUD dataset for ensembling and parameters tuning</li> </ul> <p>Unfortunately, I could see MHS dataset too late to be able to use it in the competition.</p> <h3>Feature Extraction</h3> <p>I used the Detoxify pretrained models as feature extractor [<em>toxicity</em>, <em>severe_toxicity</em>, <em>identity_attack</em>…] because <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/303665" target="_blank">they work very fine on previuos Jigsaw competitions</a> ,<br> I also generated other tf-idf models on JC/JUC dataset for each specific target [<em>toxicity</em>, <em>severe_toxicity</em>, <em>identity_attack</em> …] for total of 46 features.</p> <p>Some models generated by the individual features extracted from the detoxify models work fine on "unseen data" VAL e RUD (accuracy > 0.80 on RUD and > 0.69 on VAL) while tf-id are weak models ( < 0.71 on RUD and < 0.66 on VAL)</p> <h3>How to combine features:</h3> <p>As first attempt I fitted models (Ridge, RF, XGB, LGB) using RUD (with RMSE as metric) and tuning the hyperparameters using VAL: accuracy of these models varies from 0.700 to 0.707 on VAL. </p> <p>In this way, however, I was not sure of the performance of model on unseen data and so I simplified a little more.</p> <p>I built a simple model by selecting features optimizing ranked mean on VAL : <br> This model contains 11 Detoxify features and 1 tf-idf feature <a href="https://www.kaggle.com/steubk/jrsotc-fe-rud-2-linear-rank-sub" target="_blank">Submission 1</a></p> <p>I further eliminated 2 features which lower the accuracy on RUD to generate another model: <br> This model contains only 10 Detoxify features <a href="https://www.kaggle.com/steubk/jrsotc-fe-rud-2-linear-rank-sub-2" target="_blank">Submission 2</a></p> <p>I preferred to lower a little accuracy on VAL hoping for a better generalization (and luckily it worked ! )</p> <table> <thead> <tr> <th>Submission</th> <th>VAL</th> <th>RUD</th> <th>LB</th> <th>Private</th> </tr> </thead> <tbody> <tr> <td>sub 1</td> <td>0.70858</td> <td>0.80888</td> <td>0.759</td> <td>0.80901</td> </tr> <tr> <td>sub 2</td> <td>0.70652</td> <td>0.82242</td> <td>0.764</td> <td>0.81299</td> </tr> </tbody> </table> <h3>Public LB/Private LB</h3> <p>This is the performance of my models:</p> <table> <thead> <tr> <th>Model</th> <th>LB</th> <th>Private</th> </tr> </thead> <tbody> <tr> <td>Ridge</td> <td>0.72559</td> <td>0.78464</td> </tr> <tr> <td>RF</td> <td>0.74823</td> <td>0.80405</td> </tr> <tr> <td>LGB</td> <td>0.75595</td> <td>0.80466</td> </tr> <tr> <td>XGB</td> <td>0.75454</td> <td>0.80159</td> </tr> <tr> <td>SUB 1</td> <td>0.75922</td> <td>0.80901</td> </tr> <tr> <td>SUB 2</td> <td>0.76412</td> <td>0.81299</td> </tr> </tbody> </table> <p>With this data someone can argue that there is some signal on LB and "trust LB" was not a bad idea ! </p> <p>This is the link of the experiments I did <a href="https://github.com/steubk/jrsotc-3rdplace" target="_blank">https://github.com/steubk/jrsotc-3rdplace</a></p> <p>Sure ! this is the link <a href="https://www.kaggle.com/steubk/jrsotc-fe-rud-2-linear-rank-sub-2" target="_blank">SUB 2</a> for the 0.81299 solution </p> <p><a href="https://www.kaggle.com/steubk/jrsotc-fe-rud-2-linear-rank-sub" target="_blank">SUB 1</a> and <a href="https://www.kaggle.com/steubk/jrsotc-fe-rud-2-linear-rank-sub-2" target="_blank">SUB 2</a> are my 2 selected submission notebooks</p> <p>Muito obrigado, Marilia !</p> <p>Grazie Riccardo !</p>
Jigsaw Rate Severity of Toxic Comments	Creativity and CV [golden medal level solution: private 0.813]	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Hello! One of my decision hits the zone of golden medals (unexpectedly to me) and I would like to introduce it to community. <br> Thanks to the organizers for an interesting task.</p> <p><strong>CV</strong><br> I've used 2 validation sets:<br> 1) "Refined" validation pairs: I've removed inconsistencies in validation dataset, just keeping only unique pairs where sum(text1 is less_toxic) < sum(text2 is more_toxic). <br> 2) "Refined" dataset + 5k random pairs from ruddit<br> Both datasets correlated with private data but did not with LB.</p> <p><strong>Targets</strong><br> The main task of this competition (I guess) was to convert different types of toxicity from Jigsaw classification and others competitions to single number. So it required some kind of creativity from competitors, which is fun!</p> <p>I've used next scheme: <br> toxic_types = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']<br> 1) Xr, Yr = Take targets from reference dataset (ruddit and sample from OffenceEval)<br> 2) X = Train dataset (Jigsaw classification) texts <br> 3) Y = Train dataset scores * weights<br> 4) Train Linear(X, Y)<br> 5) Yhat = Predict (Xr)<br> 6) argmin (weights) MSE(Yhat, Yr)<br> The best weights give us the best scores<br> I've used 1 variant from public notebook, 2 variants from Ruddit and 2 variants from offenseEval, total 5</p> <p>example of code:</p> <blockquote> <p>r = pd.read_csv("data/ext_ruddit.csv")<br> c = pd.read_csv("data/jigsaw_clf.csv")<br> r = preprocess(r, 'comment_text')<br> c = c.query("labeled==True")</p> <h4>clf downsampling</h4> <p>s = c[cols].sum(1)<br> cd = pd.concat((c[s!=0], c[s==0].sample(c[s!=0].shape[0], random_state=42)))<br> cd = preprocess(cd, 'comment_text')</p> <h4>train linear model (clf texts)</h4> <p>x = cd[cols].values<br> vec = TfidfVectorizer(analyzer='char_wb', max_df=0.8, min_df=1, ngram_range=(2, 5))<br> X = vec.fit_transform(cd['comment_text'])<br> rx = vec.transform(r['comment_text'])<br> ry = r.score.values<br> w = np.array([some numbers])<br> y = x @ w<br> model = Ridge(al)<br> model.fit(X, y)<br> rp = model.predict(rx)<br> print(mse(ry, rp))> </p> </blockquote> <p>Unfortunately, I didn't score cv for ensemble of this solution (just forgot, I believe) and I used worse ensemble (with single score) of models for final solutions.</p> <p><strong>Conclusions</strong></p> <ul> <li>Transfer knowledge from different datasets could be <em>very useful</em> (aka multitask learning), but if one's done it right</li> <li>Trust your cv</li> </ul> <p><strong>Inference notebook</strong><br> <a href="https://www.kaggle.com/alturutin/inference-private-0-8128" target="_blank">https://www.kaggle.com/alturutin/inference-private-0-8128</a></p> <p><strong>Train</strong><br> U could find example of training in <strong>jigsaw22models-infer/train_regression.py</strong><br> In short:</p> <ul> <li>Backbone: <a href="https://huggingface.co/SkolkovoInstitute/roberta_toxicity_classifier" target="_blank">https://huggingface.co/SkolkovoInstitute/roberta_toxicity_classifier</a> (patriotic choice 🙂)</li> <li>Seqlen: 128</li> <li>Loss: MSE</li> <li>Gradually unfreezing head, than layers 9-12, 5-8, all</li> <li>Head = concat([CLS, GlobalPoolingMax, GlobalPoolingAvg])</li> </ul> <p>Good luck everyone in future competitions and happy kaggling :)</p>
Jigsaw Rate Severity of Toxic Comments	Customed loss(10th solution)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>There are many no-working things before I come to my final solution, the list could build an unbalanced dataset together with the working things😄<br> I always believe in simpleness. Everystep should be taken with a reason or never takes it.</p> <h5>Modeling</h5> <p>In the end, I ensembed 7 of the variation models along:<br> <strong>3 pretrain encoder</strong> roberta-base, roberta tweet sentiment, roberta tweet offensive<br> <strong>3 head</strong> bi-lstm with layernorm, robertaforclassification, robertaforclassification+sigmoid<br> <strong>4 dataset</strong> jigsaw toxic, ruddit, davidson, olid<br> <strong>2 loss</strong> customed loss and bce loss<br> <strong>seeds</strong> when the dataset keeps , change the seed. When the model keeps, change the seed. When loss keeps, don't change the seed.<br> scores are summed simplely.<br> <em>There is one submission without the oild dataset could score 5th. So the olid is not good</em></p> <h5>Customed loss</h5> <p>The loss is something like a margin rank loss. Just adapt it to be batch-wised. <br> Sort the batch samples by label score ascendingly. Sort the output in the same order. And calculate the rank loss between adjacent samples. When the batch is N samples, there are N-1 pairs is summed and averaged.<br> As well as customed loss, it benefits to have bigger batch size. I train with tpu 8 core, the batch size is 48*8=384. I kinds of assume that a best tranning is the one without strange lr schedulers, use a constant 1e-4.</p> <h5>Trust cv</h5> <p>As soon as I realized the public lb is a misleading one. I closed my eyes to it. I use both validation data and its voting version with 1-fold. And remove leaky comments in toxic data.</p> <p>It's actually a boost version of the margin loss which compares the toxicity of 2 different comments. I want to make the model be capable of comparing more than just 2 comments in a single training batch. This loss sorts all the samples in one batch by their toxicity scores.</p> <p>There is anotherfundemental sense here. For a transformer-based model，it's very important to teach it an intact information flow in very a single batch. That's very different from a CV model like Resnet, which you can teach it about the flowers in 1st batch and birds in 2nd batch. Languages are much more abstract information than images. One must carefully design the trainning pipeline to train a bert model. So I want to teach the model in every single batch whole the information in the trainning set. That's kind of a normalization operation, just to trainning samples here.</p>
Jigsaw Rate Severity of Toxic Comments	Detoxify Baseline ended up in the silver zone!	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Increasing max_len to 500 slightly increases the performance, the private lb score change from 0.80857 to 0.80868</p> <p>The reason why out-of-the-box detoxify worked so well is clearly explained by <a href="https://www.kaggle.com/readoc" target="_blank">@readoc</a> in this discussion thread <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/306181" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/306181</a> </p> <p>Somehow the Detoxify model scored extremely well in the private LB (0.80857), which is in the silver zone. So we don't really need to do anything fancy in order to get a really high score!</p> <p><a href="https://www.kaggle.com/bachngoh/using-detoxify-in-offline-mode-0-808-private-lb" target="_blank">https://www.kaggle.com/bachngoh/using-detoxify-in-offline-mode-0-808-private-lb</a></p>
Jigsaw Rate Severity of Toxic Comments	Forgotten Silver Medal	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Hello everyone, I decided to write short story about how I got my first medal.<br> I saw this competition just right after its launch and decided to join it, but after couple tries I got that this competition is fully random and here I made a goal for myself - create stable solution only trusting my CV. So, I just fine-tuned RoBERTa-base on <a href="https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset" target="_blank">Ruddit</a> and <a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge" target="_blank">Jigsaw</a> datasets and then averaged models predictions. And after this solution I stopped solving this competition. 2 months later I got an email from Kaggle with deadline approaching things and I decided to choose best by LB and best by CV solutions and thats it. I even didn't expect to get any medals, but here I am. Really lazy medal</p> <p>P.S. Honestly, It was really difficult to remember which was my solution, cuz I even have no training scripts now</p>
Jigsaw Rate Severity of Toxic Comments	Learning To Rank (#26)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Congrats to all!<br> Thanks to the host for this challenge but also for the work you share.</p> <p>I'm hesitant to post this given the lack of interest in my previous post. But it is worth a try if only one can get something out of it. </p> <h1>Framing the problem</h1> <p>What is the task? I understood the task as a <strong>ranking task</strong>. We have to rank/order comments by their toxicity importance.<br> LTR (Learning To Rank) is an interesting field. <a href="https://everdark.github.io/k9/notebooks/ml/learning_to_rank/learning_to_rank.html" target="_blank">https://everdark.github.io/k9/notebooks/ml/learning_to_rank/learning_to_rank.html</a> is a great resource to begin with this new field for me. <a href="https://embracingtherandom.com/machine-learning/tensorflow/ranking/deep-learning/learning-to-rank-part-1/" target="_blank">https://embracingtherandom.com/machine-learning/tensorflow/ranking/deep-learning/learning-to-rank-part-1/</a> + <a href="https://embracingtherandom.com/machine-learning/tensorflow/ranking/deep-learning/learning-to-rank-part-2/" target="_blank">https://embracingtherandom.com/machine-learning/tensorflow/ranking/deep-learning/learning-to-rank-part-2/</a> also help. The field seems to have grown over <a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/MSR-TR-2010-82.pdf" target="_blank">https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/MSR-TR-2010-82.pdf</a>.</p> <h1>Modeling</h1> <p>All the models in the ensemble I built are based on a personal implementation of <strong>TFR-BERT</strong> using TFR losses. My implementation would rather be TFR-xBERTx because it can be used with different backbones. Resources: <a href="https://arxiv.org/pdf/1812.00073.pdf" target="_blank">https://arxiv.org/pdf/1812.00073.pdf</a>, <a href="https://www.tensorflow.org/ranking" target="_blank">https://www.tensorflow.org/ranking</a>, <a href="https://github.com/tensorflow/ranking" target="_blank">https://github.com/tensorflow/ranking</a>, <a href="http://bendersky.github.io/res/TF-Ranking-ICTIR-2019.pdf" target="_blank">http://bendersky.github.io/res/TF-Ranking-ICTIR-2019.pdf</a>.<br> <a href="https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-2007-40.pdf" target="_blank">https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-2007-40.pdf</a> shows that pairwise approach can be improved by a <strong>listwise approach</strong> (see also <a href="https://arxiv.org/pdf/2001.01828.pdf)" target="_blank">https://arxiv.org/pdf/2001.01828.pdf)</a>. The best models I built optimize a listwise loss.<br> The idea is simple and seems to be behind Gmail search for example. A training data example is a concatenation between a query and a resource. The targets are the annotations between for example 0 and 5 concerning the relevance (I made it with 11 levels, 0 to 10, because of the 10 annotators for each comment in the training data I built). The algorithm learn by extracting the more relevant resources from a list or by ordering it. I implemented it by removing the query assuming definition of toxicity is latent but I wish I could try by giving the exact task asked to the annotators. This could allow to use all the datasets but also to let the algorithm extract latent meanings more broadly. This could make a more generic solution.<br> My solution is heavily inspired by <a href="https://arxiv.org/pdf/2004.08476.pdf" target="_blank">https://arxiv.org/pdf/2004.08476.pdf</a> and by the correponding LB: <a href="https://microsoft.github.io/MSMARCO-Passage-Ranking-Submissions/leaderboard/" target="_blank">https://microsoft.github.io/MSMARCO-Passage-Ranking-Submissions/leaderboard/</a>. It is a blend of 7 BERT-based models : RoBERTa, Deberta and Electra pointwise approach (optimize cross-entropy loss) + RoBERTa optimizing PairwiseLogisticLoss (<a href="https://www.tensorflow.org/ranking/api_docs/python/tfr/keras/losses/PairwiseLogisticLoss" target="_blank">https://www.tensorflow.org/ranking/api_docs/python/tfr/keras/losses/PairwiseLogisticLoss</a>) + RoBERTa, Deberta and Electra optimizing a listwise loss (<a href="https://www.tensorflow.org/ranking/api_docs/python/tfr/keras/losses/SigmoidCrossEntropyLoss" target="_blank">https://www.tensorflow.org/ranking/api_docs/python/tfr/keras/losses/SigmoidCrossEntropyLoss</a> worked best). The blending also includes 2 LightGBM models (<a href="https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRanker.html" target="_blank">https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRanker.html</a>) that optimize LambdaMART on a concatenation of pointwise and listwise embeddings from the 3 models of each.</p> <h1>Data</h1> <p>Exploring all the datasets shared (and more: <a href="https://hatespeechdata.com/" target="_blank">https://hatespeechdata.com/</a> + <a href="https://www.turing.ac.uk/research/research-programmes/public-policy/online-hate-research-hub" target="_blank">https://www.turing.ac.uk/research/research-programmes/public-policy/online-hate-research-hub</a>) shows that the latent definition of toxicity is very important. In my experimentations, the validation data respond better to the datasets from previous challenges from Jigsaw. It is maybe due to the instructions that the annotators have been asked. As a former psychologist I know that it is crucial and that it is a major source of bias. Jigsaw papers add to this point of view: <a href="https://storage.googleapis.com/pub-tools-public-publication-data/pdf/66073ca7ac60ee38e93fc1d173a09cab65f2fef3.pdf" target="_blank">https://storage.googleapis.com/pub-tools-public-publication-data/pdf/66073ca7ac60ee38e93fc1d173a09cab65f2fef3.pdf</a>, <a href="https://raw.githubusercontent.com/conversationai/unintended-ml-bias-analysis/main/archive/presentations/measuring-mitigating-unintended-bias-paper.pdf" target="_blank">https://raw.githubusercontent.com/conversationai/unintended-ml-bias-analysis/main/archive/presentations/measuring-mitigating-unintended-bias-paper.pdf</a>, <a href="https://aclanthology.org/2021.hcinlp-1.3.pdf" target="_blank">https://aclanthology.org/2021.hcinlp-1.3.pdf</a> as well as Ruddit paper: <a href="https://arxiv.org/pdf/2106.05664v2.pdf" target="_blank">https://arxiv.org/pdf/2106.05664v2.pdf</a>.<br> As said in the description of the competition, comments in validation_data.csv come from the Wikipedia corpus (and not only one half), more precisely the 1st challenge extract from it. EDA shows that the sampling were done by selecting ~40% of the comments that have been annotated to 1 to at least one of the targets. Only ~2.5% of the "all zero" have been sampled. This means that Jigsaw aims at ranking the more toxics one, so no need to add too much non-toxic comments in the training dataset.<br> <a href="https://arxiv.org/pdf/1610.08914.pdf" target="_blank">https://arxiv.org/pdf/1610.08914.pdf</a> underlines that optimizing cross-entropy for toxicity ratio (and not a binary task) improve predictions. I validate that. I reconstructed the dataset from the 1st Jigsaw challenge (<a href="https://meta.wikimedia.org/wiki/Research:Detox" target="_blank">https://meta.wikimedia.org/wiki/Research:Detox</a>) with 10 annotators. This was tricky because a lot of processing has been made between the initial dataset and the dataset given for the competition. Reverse engineering the data of the second challenge shows that there are at least 4 types/sets of comments (by number of annotators for a comment): I kept the comments annotated by more than 4 annotators (i.e. toxic comments). The final dataset I used for training is composed by <strong>500K+ comments annotated by at least 10 annotators</strong>. The target is the ratio of toxicity (0.2 for a comment evaluated toxic by 2 annotators on 10).<br> I have tried a lot of processing based on psychological concepts, for example the relation to norm. This did not degrade but not improve the results either. <a href="https://homes.cs.washington.edu/~msap/social-bias-frames/" target="_blank">https://homes.cs.washington.edu/~msap/social-bias-frames/</a> is a great entrypoint.<br> Trying to find a good data ordering (based on <a href="https://arxiv.org/pdf/2002.06305.pdf" target="_blank">https://arxiv.org/pdf/2002.06305.pdf</a>) failed.</p> <h1>Validation</h1> <p>I used as CV the whole <strong>CV</strong> set (0.7092/0.7602) + verification that the trend was the same on 2 independent subsets + <strong>Ruddit</strong> pairs (0.715+, scoring method: <a href="https://www.kaggle.com/sorenj/scoring-ruddit-comments" target="_blank">https://www.kaggle.com/sorenj/scoring-ruddit-comments</a>) + <strong>Measuring Hate Speech</strong> dataset (0.740+, scoring method by reproducing majority aggregation Average Agreement with Annotators on ~160,000 random pairs, 4 pairs / comment).<br> I used 4 benchmarks (3 Detoxify models + PerspectiveAPI) which score from 0.693 to 0.698 on the CV, 0.68 to 0.695 on Ruddit and 0.722 to 0.735 on the Measuring Hate Speech dataset.<br> My goal was to build a solution that could be used in a distillation process.</p> <h1>Machine learning and machine teaching</h1> <p>Sometimes I think that I do not practice ML but machine teaching. Choosing the data to give to the algorithm for example. Some talk about data-centric vs model-centric approach (see Andrew NG). It turns out that there are resources for this paradigm: <a href="https://pages.cs.wisc.edu/~jerryzhu/machineteaching/" target="_blank">https://pages.cs.wisc.edu/~jerryzhu/machineteaching/</a> and <a href="https://arxiv.org/pdf/1707.06742.pdf" target="_blank">https://arxiv.org/pdf/1707.06742.pdf</a> for example.</p> <h1>Mentoring</h1> <p>I would love if Kaggle could offer a mentorship program.</p> <p><a href="https://www.kaggle.com/cpmpml" target="_blank">@cpmpml</a>: thank you so much for your contributions to the community. I feel so much grateful! I come from very far (former psychologist) and you play(ed) an important role in my learnings for applying ML. I am sure that I am not alone. Please continue!</p> <h1>Opening</h1> <p>NLP is a very interesting field. As a psychologist I thought that language underpin abstraction and this is why I learn it. There is so much to build and to learn.</p> <p>Thanks if you read until here.</p> <p>Thanks!<br> Congrats <a href="https://www.kaggle.com/amontgomerie" target="_blank">@amontgomerie</a> for the solo gold! You deserve it. Your interventions were constructive and interesting.</p>
Jigsaw Rate Severity of Toxic Comments	My First Silver Medal: Solution Approach (Simple & Single model)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p><strong>Step-0</strong>: Basic Text Cleaning<br> <strong>Step-1</strong>: Used Data from <a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge</a> and BERT ('bert-base-uncased') model to train to predict probabilities for different classes. <br> <strong>Step-2</strong>: Trained model is used to predict class-probabilities for the comments for this competition (Train/validation Data)<br> <strong>Step-3</strong>: Used Bayesian Optimization to find weights for each of the classes such that the class-probabilities when multiplied with the weights optimize the validation data shared in this competition<br> <strong>Step-4</strong>: Submission: Load the Trained BERT model, predict class-probabilities, and multiply with the optimized weights to predict score, followed by ranking to avoid duplicate score.</p> <p><strong>Scope for improvement (Simple Ones)</strong>: Multi-folds / Additional Data-Sets / Ensemble-Methods / Non-Linear weight multiplier across different classes (instead of direct sum(w*c))</p> <p><strong>Thanks to Kaggle and other experts for sharing their valuable insights/experiences throughout the competitions. Special thanks to all those who helped me realize (mathematically) that score more than ~0.82+ is not real</strong></p>
Jigsaw Rate Severity of Toxic Comments	Pretrained model benchmark + 0.769 Private LB solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>As i mentioned earlier, here is the <strong>table for toxicity classification benchmark with offline datasets, taken from huggingface.co ready to use as a pre-trained models</strong>.</p> <p>I hope it will be usefull for somebody!</p> <table> <thead> <tr> <th>Model</th> <th>LB</th> <th>Offline Dataset</th> </tr> </thead> <tbody> <tr> <td>twitter-roberta-base-offensive</td> <td>0.727</td> <td><a href="https://www.kaggle.com/dzisandy/twitterrobertabaseoffensive" target="_blank">link</a></td> </tr> <tr> <td>bert-base-uncased-hatexplain-rationale-two</td> <td>0.784</td> <td><a href="https://www.kaggle.com/dzisandy/bertbaseuncasedhatexplainrationaletwo" target="_blank">link</a></td> </tr> <tr> <td>SkolkovoInstitute/roberta_toxicity_classifier_v1</td> <td>0.743</td> <td><a href="https://www.kaggle.com/dzisandy/skolkovoinstituteroberta-toxicity-classifier-v1" target="_blank">link</a></td> </tr> <tr> <td>abhishek/autonlp-toxic-new-30516963</td> <td>0.750</td> <td><a href="https://www.kaggle.com/abhishek/autonlp-toxic-1" target="_blank">link</a></td> </tr> <tr> <td>twitter-roberta-base-hate</td> <td>0.716</td> <td><a href="https://www.kaggle.com/dzisandy/twitterrobertabasehate" target="_blank">link</a></td> </tr> <tr> <td>dehatebert-mono-english</td> <td>0.728</td> <td><a href="https://www.kaggle.com/tunminhhunh/hatespeechcnergdehatebertmonoenglish" target="_blank">link</a></td> </tr> <tr> <td>twitter-roberta-base-sentiment</td> <td>0.708</td> <td><a href="https://www.kaggle.com/dzisandy/twitterrobertabasesentiment" target="_blank">link</a></td> </tr> <tr> <td>SkolkovoInstitute/roberta_toxicity_classifier</td> <td>0.654</td> <td><a href="https://www.kaggle.com/jonathanchan/roberta-toxicity-classifier" target="_blank">link</a></td> </tr> <tr> <td>twitter-xlm-roberta-base-sentiment</td> <td>0.647</td> <td><a href="https://www.kaggle.com/dzisandy/twitterxlmrobertabasesentiment" target="_blank">link</a></td> </tr> <tr> <td>distilbert-base-uncased-finetuned-sst-2-english</td> <td>0.625</td> <td><a href="https://www.kaggle.com/dzisandy/distilbertbaseuncasedfinetunedsst2english" target="_blank">link</a></td> </tr> <tr> <td>distilroberta-finetuned-tweets-hate-speech</td> <td>0.608</td> <td><a href="https://www.kaggle.com/dzisandy/distilrobertafinetunedtweetshatespeech" target="_blank">link</a></td> </tr> <tr> <td>unbiased-toxic-roberta</td> <td>0.535</td> <td><a href="https://www.kaggle.com/dzisandy/unbiased-toxic-roberta" target="_blank">link</a></td> </tr> <tr> <td>unitary/toxic-bert</td> <td>0.467</td> <td><a href="https://www.kaggle.com/julian3833/toxic-bert" target="_blank">link</a></td> </tr> </tbody> </table> <p>All model are taken from huggingface.co, datasets are ready to use as:<br> <code>model_path = '/path to dataset/ '</code><br> <code>tokenizer = AutoTokenizer.from_pretrained(model_path)</code><br> <code>model = AutoModelForSequenceClassification.from_pretrained(model_path)</code></p> <p>Using this models, we used to obtain best result with, as a blending rule Ridge Regression is used:</p> <table> <thead> <tr> <th>Public LB</th> <th>Private LB</th> </tr> </thead> <tbody> <tr> <td>0.78959</td> <td>0.79641</td> </tr> </tbody> </table> <p>Unfortunately, we haven't trusted this model and haven't used it for a submission. Very unlucky :(<br> <strong>But you can watch this solution <a href="https://www.kaggle.com/dzisandy/easy-bert-based-models-blending-private-0-79641" target="_blank">here</a></strong></p> <p>Congratulations on winning prizes! Happy kaggling! If you use the data provided here, please upvote!</p>
Jigsaw Rate Severity of Toxic Comments	Private score 0.81148. What could've been the 5th place solution (Single model)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Congratulations to all winners. I wasn't able to focus on the competition due to some other commitments during the latter half. During the start, I had my eye on one of my baseline models even though it didn't perform well in public LB. It ended up scoring 0.81148 in private LB. So I thought I'll share the approach anyway. </p> <h2>Data prep</h2> <h3>Evaluation data</h3> <p>I used the text pairs provided as validation data, but converted them into a regression score by fitting the pairs to a Bradley-Terry model. <a href="https://www.kaggle.com/jdoesv/wrangled-validation-data" target="_blank">Notebook</a></p> <h3>Training data</h3> <p>Due to the severe overlap between this competition's data and Jigsaw 1's data, I used that original competition's dataset here. All samples that were toxic in anyway (along with arbitrarily sampled non-toxic texts) were used. The weights for each of the six columns in the original dataset are the inverse ratio of their counts. That way, all types of toxic data were normalized to the same plane. I also standardized them to fit between 0. and 1. <br> <a href="https://www.kaggle.com/jdoesv/baseline-data-prep/notebook#Train-data" target="_blank">Notebook available here</a></p> <h3>Data preprocessing for both train and validation</h3> <p><a href="https://www.kaggle.com/jdoesv/wt-text-processing-utils/notebook" target="_blank">Notebook</a><br> Preprocessing was fairly simple and done mostly using regex</p> <ol> <li>Removed IP Addresses</li> <li>De-emojisized all emojis.</li> <li>Unidecoding.</li> <li>All repetitions were reduced to a maximum of 3 repetitions. E.g. : <code>hellllllllllllllllo</code> -> <code>helllo</code>, and <code><swearword> <swearword> <swearword> <swearword> <swearword>.....</code> -> <code><swearword> <swearword> <swearword></code></li> </ol> <h2>Model Selection:</h2> <p>One intuition that I had was to use a model <strong><em>pretrained</em></strong> on conversation style corpora (which are two-way communications) instead of the more stable Bert-like models which are usually pretrained on Wiki, Bookcorpus and similar one-way communication texts which are polished over and over, unlike comments that are made on the fly (like this competition's dataset). The only model that came close to this requirement was <a href="https://huggingface.co/vinai/bertweet-large" target="_blank">Bertweet Large</a> . I finetuned it for <strong><em>pointwise ranking</em></strong> regression. <br> <a href="https://www.kaggle.com/jdoesv/baseline-model-bertweet-regression" target="_blank">Training</a> and <a href="https://www.kaggle.com/jdoesv/baseline-predictions/notebook?scriptVersionId=80303997" target="_blank">inference</a></p> <p>Note to self: Bookmark competition end dates in calendars so I can allocate some time selecting the final submissions. </p> <blockquote> <p>Did you also use jigsaw1 in your selected models?</p> </blockquote> <p>I used both Jigsaw 1 and 3 (Unintended bias) datasets. </p> <blockquote> <p>why didn't you select it?</p> </blockquote> <p>Terrible oversight. I thought the competition was ending today and hence didn't select my two candidate models.</p>
Jigsaw Rate Severity of Toxic Comments	Single Bert [1 epoch] to get silver medal	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><ul> <li><strong>pretrained model</strong>: <a href="https://huggingface.co/GroNLP/hateBERT" target="_blank">HateBert</a></li> <li><strong>task type</strong>: regression</li> <li><strong>score</strong>: 0.80077(1 epoch), 0.80345(5 epochs)</li> <li><strong>code</strong>: <a href="https://www.kaggle.com/leolu1998/jigsaw-hatebert-reg-infer" target="_blank">Jigsaw - HateBert Reg Infer</a></li> </ul>
Jigsaw Rate Severity of Toxic Comments	Toxic Solution and Review (2nd Place)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thanks for your praise, maybe the green hand would more straight-forword and directly put the knowledge into practice LOL. </p> <p>Thanks! I suppose that the notebook had been public. The loss function I used is MSE by <code>torch.nn.MSELoss()</code></p> <p>First of all, congrats to the winners and everyone who enjoyed the competition! ! Although this is my first time participating in a competition, I really like this competition on Kaggle, especially the competitive atomsphere in there.</p> <h1>Overview</h1> <p>To be honest I never thought of finishing in 2nd place, which is largely due to the luck, but I will say ''Thank you!'' to the luck for seeing positive results for my persistence with <strong>''Bert''</strong> Method.I have to say that we could marvel at and believe the performance of "Bert" in NLP world all the time. The Roberta helps me attain the success, even though its' public score is lower than linear model, but I find that the pinpoint of this competition perhaps is that ''Public leaderboard perhaps mislead us to be overfitting'' </p> <h1>Retrospective</h1> <p>What I remember fresh is probably in the first ten days of January, when many people started to increase the public leaderboard ranking crazily, which made me feel that the overall trend was not normal, so I made two decisions: </p> <ol> <li>Running program Locally in order to avoid paying more attention to the public leaderboard scores.</li> <li>I chose to part ways with a partner who was supposed to be teaming up. (The story here is that we communicated and did a lot of groundwork and exploration togethera at the beginning, but in those crazy period in January, I disagree with his obsession with improving public scoring and trust linear model instead of BERT, (Perharps the submission score of Roberta only 0.79416, which is even lower than 0.8 and not to say the linear model(0.893+)), but it would cause some overfitting risks I realized when I learn more from papers and my professor. <br> Obviously, that's the reason why we can see this remarkable 'Shake-up' from Public leaderboard to Private Leaderboard. Fortunately, the Luck resonate my persistence in BERT method. </li> </ol> <h1>About Dataset</h1> <p>I used data from the last toxic classification competition. Since the goal of this competition is to predict the toxicity of comments, which is a regression task, I weighted averaged several features of the orginal data.</p> <h1>About Data Cleaning</h1> <p>BERT uses raw data when pretrained, which means that Pre-trained BERT is based on an uncleaned corpus. Actually, I find that Data Cleaning will change distribution and lose information. <br> Usually, data cleaning can improve the performance of linear models, but could harm the performance of BERT. If the weight decay or dropout value is too large, the model performance will degrade. When we go deep, we could find that Linear models need to be cleaned to reduce feature size and reduce overfitting. The capacity of BERT is large enough that it is not very necessary.<br> So in the end, I make a huge change of the code, especially simpliying the data cleaning procedure</p> <p>Comparison：<br> Apply Data cleaning： Private Score：0.80826， Public Score：0.77609<br> Delete some Data cleaning： Private Score：0.81363， Public Score：0.79416</p> <h1>About Feature Engineering</h1> <ol> <li>For text preprocessing, only simple tokenization methods are used. I didn't select the features manually because the model can learn the importance of each token.<br> 2.The BERT model acts like a feature transformer, converting text data into a vector representation of length 768.</li> </ol> <h1>About Train Method</h1> <p>I used BERT(Bidirectional Encoder Representations from Transformers) for this competition. BERT is currently the State-of-the-art language model that makes use of Transformer to learn contextual relations between words (or sub-words) in a text. The model I used in this competition consists of two parts: the RoBERTa base, and a multi-layer perceptron built upon it for regression. That‘s is the essence of idea why I choose this method and insist it until the competition is ended.</p> <h1>About Esemble Weight</h1> <p>I randomly sampled 5 folds of data, each fold contained a proportion of the data and trained a separate Model on each fold. Then average up the predictions from the 5 models.</p> <h1>Optimization</h1> <p>I used Adam optimizer as it is computationally efficient and works well with large data sets and large parameters. Besides, I also used a learning rate schedular to decrease the learning rate gradually based on the epoch. This has the effect of quickly learning good weights early and fine tuning them later.</p> <h1>Avoid Overfitting</h1> <p>Focusing on using Weight Decay, Dropout Layers and Early stopping to prevent the models from overfitting. All of these keep the relative good generalization of model.<br> To be more specific, For <em>Earlystoping</em>, the stopping criteria will be based on the validation loss in each epoch.</p> <p>Thanks everyone，I am happy everyone help new bird , I learn a lot from this competition! </p> <p>I‘m happy that it could encourage you a lot ! From my perspective, Less overfitting exert an essential role in competition, I suggest you could review the knowledge about it, keep the model having better generalization. Besides, you also directly participate the competition of amateur such as" Titanic", which provide specifc guide for you, which means that it is a new bird friendly competition. And, Just do it !!!! You could never believe something that would happend from public leaderboard to private leaderboard.</p>
Jigsaw Rate Severity of Toxic Comments	Trust your CV?	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>What's your CV schema? It seems to be important to choose right CV there.<br> Mine is stratified k-fold by connectivity components of aggregated validation data by majority voting.</p> <pre><code>g = nx.Graph() for _, row in df_majority.iterrows(): less_toxic = row['less_toxic_id'] more_toxic = row['more_toxic_id'] g.add_edge(less_toxic, more_toxic) sizes = [] groups = [] for c in nx.connected_components(g): sizes.append(len(c)) groups.append(c) for ..., ... StratifiedKFold().split(..., y=sizes): # set folds to df_majority pairs </code></pre> <p>Also I used <strong>ruddit pairs for validation.</strong> Well, final solution is MLP-stacking with 818LB, 809PLB and 10 fold 760+ CC validation and 710+ ruddit validation. Feel free to share your CV schema or ideas to choose your final solution ;)</p>
Jigsaw Rate Severity of Toxic Comments	Unselected top scored simple Kmeans solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Simple kmeans clustering model and Euclidean distance from centroid of non-toxic cluster scored 0.818 in private LB. </p> <p><img src="https://i.postimg.cc/hjkKhC0K/Capture.jpg" alt=""></p> <p>K means clustering was performed on Jigsaw Multilingual Toxic Comment Classification dataset and validation data from this competition untouched. It scored 0.718 in validation data of this competition.</p>
Jigsaw Rate Severity of Toxic Comments	[45th solution]Lucky Beginner's only 1 submission	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h1>First</h1> <p>A kaggle beginner got a silver medal by only 1 submission with "great luck & a few ideas", so here is the solution<br> It is not about advanced strategies or difficult techniques, and there may be some technical errors and poor English.</p> <h1>Overview</h1> <p>Public LB is unreliable, so I used validation_data.csv, which is the dataset of this competition, as the evaluation data.<br> However, the validation_data also contains noise that was randomly selected by the annotator.<br> I thought that the annotator might randomly select the data in the following patterns, so I deleted the data in the following patterns in validation_data and used it as an evaluation data set. The details are posted <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/303429" target="_blank">here</a>.<br> (1) data with similar toxicity scores, <br> (2) data with lower toxicity scores, <br> (3) data with similar meanings (By sentence transformers. Code is <a href="https://www.kaggle.com/mujrush/reliable-validation-data-by-sentence-transformers" target="_blank">here</a>)</p> <h1>Procedure</h1> <ul> <li><p>Step 1. Model training and prediction<br> Using the first jigsaw competition data, train some models to predict the toxicity score of validation_data.</p></li> <li><p>Step 2. Pre-processing of validation_data<br> Remove the data corresponding to steps 1-3 from the prediction data of each top model in step 1. As a result, the validation_data has been reduced to half the size.<br> (1) Remove data with a close toxicity score (+0.01 to -0.01)<br> (2) Remove data with low toxicity score (0.15 or less) <br> (3) Remove data with close meaning (0.65 or higher by sentence transformers)</p></li> <li><p>Step3. Model evaluation by pre-processed validation_data<br> Predict with some models again with pre-processed validation_data from step2 and submit the best score in the ensemble.</p></li> <li><p>(Step4. Pray to get a medal. This procedure is not reproducible)</p></li> </ul> <h1>Dataset</h1> <p>The data from <a href="https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge" target="_blank">the first jigsaw competition</a></p> <h1>Model</h1> <p>1.BERT-jigsaw-severetoxic<br> 2.roberta<br> 3.hatebert<br> 4.berta <br> 5.twitter-roberta-base-offensive<br> 6.TFIDF+ridge</p> <h1>Traing</h1> <ul> <li>fold :5</li> <li>epoch: 10</li> <li>Learning rate change: No</li> <li>Scheduler: No</li> </ul> <h1>Ensemble</h1> <p>roberta + TFIDF/Ridge</p> <p>Here's a table of model results.</p> <table> <thead> <tr> <th>model</th> <th>valid_data score</th> <th>Preprocess valid_data</th> <th>Preprocessed valid_data score</th> <th>ensemble</th> </tr> </thead> <tbody> <tr> <td>BERT-jigsaw-severetoxic</td> <td>0.691078</td> <td>yes</td> <td>0.78536680</td> <td></td> </tr> <tr> <td>roberta</td> <td>0.687491</td> <td>yes</td> <td>0.78744243</td> <td>yes(weight:0.892)</td> </tr> <tr> <td>hatebert</td> <td>0.686561</td> <td>yes</td> <td>0.78348576</td> <td></td> </tr> <tr> <td>BERT</td> <td>0.685366</td> <td>yes</td> <td>0.78478303</td> <td></td> </tr> <tr> <td>twitter-roberta-base-offensive</td> <td>0.685000</td> <td>yes</td> <td>0.78536680</td> <td></td> </tr> <tr> <td>TFIDF/Ridge</td> <td>0.681513</td> <td></td> <td>0.78108581</td> <td>yes(weight:0.108)</td> </tr> </tbody> </table> <h1>Improvements</h1> <ul> <li>Change training dataset/mode/fold/epoch/ in step1.</li> <li>Add training Learning rate change/Scheduler in step1.</li> <li>Tuning of each threshold in step2.</li> <li>Ensemble selection in step3. etc.</li> </ul> <h1>Finally</h1> <p>In this competition, I was happy to get upvotes and reactions on my first submission to kaggle kernel and discussion. I think it's great that the kaggle community is so warm and friendly, even for beginner like me.</p> <p>I hope this post will be useful for dear kaggler.</p>
Jigsaw Rate Severity of Toxic Comments	[9th solution] Ensemble of 5 external dataset transformers	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>First of all huge thanks to the organizers of the competition and to all the participants. I've learned a lot and strengthened my faith in the "Trust your CV" principle 🙂</p> <h4>High level description</h4> <p>List of models used in the ensemble:</p> <ol> <li>CV: 0.7356, LB: 0.786 - roberta-base trained on CCC 2017 dataset using custom attention head, cross-entropy and margin ranking losses, a multi-staged prediction (first predict toxicity without knowing the 6 labels, then aggregate that prediction and 6 labels into the final prediction).</li> <li>CV: 0.7450, LB: 0.747 - roberta-base trained on CCC 2017 dataset using a weighted average of margin ranking and binary cross-entropy loss.</li> <li>CV: 0.7403, LB: 0.790 - roberta-base regression model trained on Ruddit dataset.</li> <li>CV: 0.7448, LB: 0.751 - unitary/unbiased-toxic-roberta fine-tuned on Ruddit dataset with 2-layer DNN as a regressor.</li> <li>CV: 0.7429, LB: never submitted - roberta-base regression trained on Wiki Talk Labels dataset.</li> </ol> <p>The ensembling has CV 0.7605, LB 0.779. All those models were ensembled using rankdata + mean.</p> <h4>Validation strategy</h4> <p>All of the models were validated on <code>validation_data.csv</code> aggregated using majority voting.<br> First I used to define a final CV score as <code>CV * 0.65 + LB * 0.35</code>. Those weights represented my personal confidence in the score value. At some point, I've noticed a huge flakiness in the public LB score, small model change (0.001 score difference on CV) led to the 0.05-0.1 change in the LB score. I hypothesized that a small number of unique comments are used either in <code>more_toxic</code> or in <code>less_toxic</code> roles. More about it here <a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/305301" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating/discussion/305301</a> . So I gave up on LB and stuck to the local CV score only.</p> <h4>Ensembling strategy</h4> <p>At first, I wanted to ensemble 14 models (trained on UBTC, Offenseval, C3, Hate Speech Measurement) instead of 5. But it in the ranking task more models do not necessarily give better score. The score distribution differs a lot across the models, the toxicity definition differs a lot across the datasets (Ruddit and Offenseval for example). To get maximum out of ensembling I've generated all the subsets of 14 models and run the validation on the results blended using rankdata + mean. This way I've identified that the highest score is reached by the subset of the 5 aforementioned models.</p> <p>I've also experimented a lot with various ensmebling strategies: taking mean across the closest scores only, writing custom comparator function and sorting the comments based on voting, ranking + artificial distribution generation (Poisson, Normal), etc. Unfortunately, everything worked slightly worse than rankdata + mean on CV.</p> <h4>General transformer tricks</h4> <p>The following tricks were used for training all the models:</p> <ol> <li>Per-layer LR, lower for backbone layers (2e-5), higher for DNN or final attention head (up to 1e-3).</li> <li>High and dynamic validation frequency: typically the model is trained for 1-2 epochs with validation at least every 10% batches. The higher the accuracy, the higher the validation frequency.</li> <li>AdamW optimizer and cosine annealing with warmup scheduler, warmup typically takes 1 validation cycle.</li> </ol> <h4>Things which didn't work</h4> <ul> <li>MLM pretraining</li> <li>Augmentations / TTA. Although I've only explored some basic ones, they seem to change the sense or the toxicity of the comment significantly, and thus worsen the score. Moreover, I've started playing with them at the very end of the competition so probably that's just a lack of time.</li> <li>unitary/toxic-bert backbone. It gives superior performance, but it has been trained on the whole CCC 2017 dataset and that's why some portion of validation_data.csv comments have been exposed to it before. I've considered it a leaky one and switched to roberta-base and unitary/unbiased-toxic-roberta.</li> </ul>
Jigsaw Rate Severity of Toxic Comments	[Simplest Gold medal ?]11th solution (Ridge + Detoxify + LightGBM)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Rate Severity of Toxic Comments <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><h1>WARNING</h1> <p>First of all, congratulations to all winners. <br> Though, really fortunately, I got the gold medal, <br> <strong>90%(or more?) of this result may be just lucky.</strong></p> <p>Therefore, my solution is maybe not so interesting,special, and beneficial to you .</p> <p>What I did are just </p> <ul> <li>search additional data on Kaggle datasets</li> <li>Use some libraries to get toxicity of text</li> <li>make features and train simple LightGBM</li> </ul> <p>If it's ok, then please read my solution (or my poem).</p> <h2>4 lines Overview</h2> <p>step1 : train ridge regressor by tf-idf of a lot of datasets <br> step2 : Toxicity prediction library(Detoxify)<br> step3 : convert validation_data.csv to rate of more toxic for train<br> step4 : Train LightGBM with these features and additional naive features</p> <h2>Overview_image</h2> <p><img src="https://pbs.twimg.com/media/FLFjKScUcAQRG02?format=jpg&name=large" alt="画像"></p> <h1>step1 train ridge regressor by tf-idf of a lot of datasets</h1> <h2>1-1 : reading ridge regression notebook</h2> <p>I started from <a href="https://www.kaggle.com/samarthagarwal23/mega-b-ridge-to-the-top-lb-0-85?scriptVersionId=80534222" target="_blank">this ridge regression notebook</a>.</p> <p>The notebook is</p> <ul> <li>To vectorize, It use tf-idf</li> <li>ridge regression with a lot of external data </li> <li>brute force weighting the ridge results <br> (It takes a lot of memory, so I can’t weighting not so many models )</li> </ul> <p>After reading and executing this notebook, I thought</p> <ul> <li>The more external data is, the more diverse prediction I can do<ul> <li>I thought diverse predictors are needed because the score is averaged by many workers. </li></ul></li> <li>Weighting the ridge results by another method<ul> <li>Finally, I chose LightGBM for optimal weighting the ridge results.</li></ul></li> </ul> <h2>1-2 : search additional data</h2> <p>The notebook contains these data to train ridge regression</p> <ul> <li>ruddit data<ul> <li><a href="https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset" target="_blank">https://www.kaggle.com/rajkumarl/ruddit-jigsaw-dataset</a></li></ul></li> <li>previous jigsaw data<ul> <li><a href="https://www.kaggle.com/julian3833/jigsaw-multilingual-toxic-comment-classification" target="_blank">https://www.kaggle.com/julian3833/jigsaw-multilingual-toxic-comment-classification</a></li> <li><a href="https://www.kaggle.com/c/jigsaw-toxic-severity-rating" target="_blank">https://www.kaggle.com/c/jigsaw-toxic-severity-rating</a></li></ul></li> </ul> <p>And I added </p> <ul> <li>wikipedia data<ul> <li><a href="https://www.kaggle.com/manishguptads/wikipedia-toxicity" target="_blank">https://www.kaggle.com/manishguptads/wikipedia-toxicity</a></li></ul></li> <li>Dynamically Generated Hate Speech Dataset<ul> <li><a href="https://www.kaggle.com/usharengaraju/dynamically-generated-hate-speech-dataset" target="_blank">https://www.kaggle.com/usharengaraju/dynamically-generated-hate-speech-dataset</a></li></ul></li> <li>Malignant Comment Classification<ul> <li><a href="https://www.kaggle.com/surekharamireddy/malignant-comment-classification" target="_blank">https://www.kaggle.com/surekharamireddy/malignant-comment-classification</a></li></ul></li> </ul> <p>I used them to train ridge regressor and train LightGBM with the regressor's output.<br> (continued on step 4)</p> <h1>step2 : Call toxicity prediction library</h1> <p>I use <a href="https://github.com/unitaryai/detoxify" target="_blank">Detoxify</a> . It’s toxicity scoring library by Bert like models trained with past zigsaw competitions.<br> And I found <a href="https://www.kaggle.com/steubk/detoxify-quick-prediction-without-internet" target="_blank">offline execution notebook</a></p> <p>This library have three kinds of pretrained model (original,unbiased,multilingual)<br> I use all of these models.</p> <p>I also use these models to convert train text to features .<br> (continued on step 4)</p> <h1>step3 : convert validation_data.csv to rate of more toxic for train</h1> <p>For train data, I use provided validation_data.csv.<br> 1 problem is how to make the target value.</p> <p>As long as I read competition overview, test data score is calculated like this↓</p> <p><img src="https://pbs.twimg.com/media/FLFk9DnUcAEtejb?format=png&name=large" alt="図2"><br> In other words, the score is the rate of more toxic. </p> <p>Then, to train, I did the same conversion to train data. I train my model to predict the score directly.</p> <h1>step4: Train LightGBM with these features and additional naive features</h1> <p>Finally, I concat all of these features,train,and predict score.<br> And I added some naive features(length of text, words count, and so on)<br> Please see the image I attached the beginning of this notebook.</p> <h1>result</h1> <p>public LB 0.79242<br> private LB 0.81062</p> <p><strong>Thank you for reading ！</strong><br> If you have any question, I'll answer (if I can …)</p>
Jigsaw Unintended Bias in Toxicity Classification	106th place solution (21 models blend)	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Unintended Bias in Toxicity Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Thank you! Yeah, the same story, pytorch helps you to dive deep into your models, but sometimes you have to pay for this forgetting the syntax :D</p> <p>Thank you very much, hope you get Expert in Competitions soon, as well as Master and Grandmaster ;)</p> <p>Thank you, my friend!</p> <p>Good day everyone! I was waiting eagerly for the results of this competition to write a short report about the solution of our team and what it meant to us😊</p> <p>You can find our final submitted notebook here: <a href="https://www.kaggle.com/blackitten13/lstm-cp-clip-nb-tw-tfidf-bert4x-gpt-test-best?scriptVersionId=16367978">https://www.kaggle.com/blackitten13/lstm-cp-clip-nb-tw-tfidf-bert4x-gpt-test-best?scriptVersionId=16367978</a></p> <h1>Our team</h1> <p>I started solving this competition about 2-3 weeks before the final deadline. I was spending all 5 my submissions per day, however, at some moment I understood that model blending shows itself to be very powerful in this competition, but I didn't have much time/resources to make many strong diverse models. So, I decided to blend with my friend <a href="https://www.kaggle.com/artkulak">Artem Kulakov</a>, a BSc CS student at HSE. I also wanted to find another person to merge with and saw that <a href="https://www.kaggle.com/lopuhin">Konstantin Lopuhin</a> started to make some progress towards solving the problem. He had a slightly worse score than me, though. So, I proposed a merge and he accepted it, and I was very happy about it - it is my first serious Kaggle competition, but I still had an honour to work with such a well-known Grandmaster!</p> <h1>Our approach</h1> <p>As I already said, we believed in blending, so we all tried to prepare strong diverse models and mix them properly. Artem and I were working on LSTM-based models, while Konstantin, as a master of transformers, designed BERT and GPT-2-based solutions. I also added some TF-IDF models, and came up with how can we make a cool blend.</p> <p>Next, I want to give more details about our methods. In this report, I tried to focus on what made our solutions unique in some ways - so I can skip something which was well-known due to public kernels, e.g. preprocessing for LSTM models - it is completely the same as it was in the baseline kernel to which I refer below. Also we didn't invent anything for the loss function.</p> <h3>Validation</h3> <p>As Konstantin correctly noticed, it was pointed that in the 2nd stage, the data will consist of 50% annotated identities and 50% not annotated identities. Therefore, we decided to use a holdout set of the same size as the test set which was stratified by <code>identity_annotator_count</code> column (you can see the code <a href="https://www.kaggle.com/blackitten13/my-validation">here</a> - where the train set was split to <code>train_df</code> and <code>val_df</code>). We decided not to use OOF predictions for several reasons: it could take too much time, we still had many models, the holdout set was relatively small (so it wouldn't hurt the model quality much) and also validation score correlated in a good way with public score. It is also worth noting that our models barely fit in the private dataset quota 😄 </p> <h3>LSTMs</h3> <p>As a baseline, we used <a href="https://www.kaggle.com/chechir/bert-lstm-rank-blender">this kernel</a>. Then, throughout the whole time of solving the competition, with 5 submissions per day, we tested many ideas. Finally, all our LSTM-based models had similar architecture:</p> <ul> <li>Embedding (non-trainable)</li> <li>SpatialDropout (0.3)</li> <li>LSTM #1 (128/256 units, bidirectional)</li> <li>LSTM #2 (128/256 units, bidirectional) - it took the result of LSTM #1 as an input</li> <li>Concatenation of 3 types of pooling: Global Average Pooling, Global Max Pooling and Softmax Pooling (e.g. see <a href="https://www.cs.tut.fi/sgn/arg/dcase2017/documents/challenge_technical_reports/DCASE2017_Salamon_192.pdf">[1]</a>)</li> <li>Concatenation of two layers - for output and auxiliary output, as in the baseline kernel. However, instead of being just a dense layer, we made it to be Dropout (0.5) -> BatchNorm -> Dense.</li> </ul> <p>In order to make our model diverse, we came up with some changes that we made in them.</p> <p><em><strong>Final models</strong></em></p> <ol> <li>[CP] (<a href="https://www.kaggle.com/blackitten13/bilstm-crawl-paragram-6-ep-0-train">https://www.kaggle.com/blackitten13/bilstm-crawl-paragram-6-ep-0-train</a>) - this model took concatenated Crawl and Paragram vectors as embeddings, which were provided <a href="https://www.kaggle.com/iezepov/gensim-embeddings-dataset">in this dataset</a>. It had 128 hidden LSTM units.</li> <li>[CP with target clip 0.05] (<a href="https://www.kaggle.com/blackitten13/bilstm-c-p-target-clip-0-05-5-ep-0-train-val?scriptVersionId=16115529">https://www.kaggle.com/blackitten13/bilstm-c-p-target-clip-0-05-5-ep-0-train-val?scriptVersionId=16115529</a>) - the same model, but with a trick - train target values from [0, 0.05] were clipped to 0.05, and the ones from [0.95, 1] were clipped to 0.95, in order to prevent too big model certainty and the impact of noise. And also here we changed number of hidden LSTM units to 256.</li> <li>[Nb] (<a href="https://www.kaggle.com/blackitten13/bilstm-numberbatch-0-train-val">https://www.kaggle.com/blackitten13/bilstm-numberbatch-0-train-val</a>) - LSTM model with Numberbatch embeddings (from the same dataset) with 256 hidden LSTM units.</li> <li>[Tw] (<a href="https://www.kaggle.com/blackitten13/bilstm-twitter-1-train-val">https://www.kaggle.com/blackitten13/bilstm-twitter-1-train-val</a>) - LSTM model with Twitter embeddings with 256 hidden LSTM units.</li> </ol> <p>Each model was trained with 3 different seeds, therefore for each model instead of one we obtained three of them. For the blend, we took the best epoch on validation for each model/seed. Thus we got 3 * 4 = 12 different LSTM models for the blend.</p> <p><em><strong>What was tried but didn't work</strong></em></p> <ul> <li>Adding convolutional layers in many different ways</li> <li>Attention</li> <li>Adding GRU cells</li> <li>Focal loss</li> </ul> <p><em><strong>A tragedy</strong></em></p> <p>As I wrote <a href="https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/discussion/97387">here</a>, we had a tragedy - we didn't make it on time to include models made by Artem completely (so for the final solution we took only models prepared by me and Konstantin 😔 ). He had some strong candidates - we calculated validation score for them and even included them in blend, but, unfortunately, we were late to put it all together because there were ~10 of them, and all with different preprocessings 😲 (however, as we calculated later, including them could give us a boost just in several places up). Still, I want to include his interesting thoughts:</p> <ul> <li>using regular expressions replacement with the patterns below gave a slight boost both to CV and LB scores: <code>replacement_patterns = [ (r'won\'t', 'will not'), (r'can\'t', 'cannot'), (r'(I\|i)\'m', 'I am'), (r'ain\'t', 'is not'), (r'(\w+)\'ll', '\\g&lt;1&gt; will'), (r'(\w+)n\'t', '\\g&lt;1&gt; not'), (r'(\w+)\'ve', '\\g&lt;1&gt; have'), (r'(\w+)\'s been', '\\g&lt;1&gt; has been'), (r'(\w+)\'s', '\\g&lt;1&gt; is'), (r'(\w+)\'re', '\\g&lt;1&gt; are'), (r'(\w+)\'d', '\\g&lt;1&gt; would') ]</code></li> <li>blending predictions from each models epochs with weights growing exponentialy (as was shared in many kernels) gave a big boost</li> <li>adding statistical features collected for each text sample also increased the score (although many people wrote it did not work)</li> <li>it was able to dramatically reduce train size by excluding easy training samples. It could have been done in the following way. Train a simple LSTM network for 1 epoch and then exclude all samples which were scored very close to zero (these samples were easy to classify even when no training was done with them)</li> </ul> <p><em>didn't work</em>:</p> <ul> <li>pseudo labeling (probably more time should have been invested in it to make it work)</li> <li>Crawl + Glove embeddings - worked worse than Crawl + Paragram</li> <li>in some recent competitions there was an advice to set embeddings layer to trainable for the last epoch. It slightly overfits, but adds more diversity to the final blend. Though this technique wasn't explored fully, so it did not work as well.</li> </ul> <h3>Transformers</h3> <p>For transformers, we took <a href="https://www.kaggle.com/yuval6967/toxic-bert-plain-vanila">Yuval Reina’s kernel</a> as a starting point, converted it to a script, added a custom batcher (more random and more efficient than in kernels), slightly increased train sequence length, increased test sequence length, which gave around 0.399 LB from a single model on one epoch and took around 4-5 h to train on 2080ti. Unfortunately, scaling it up was not successful - we didn’t get a better score from 2 epochs, or from bert-large. Still we added bert-base-cased and bert-large as they provided diversity in the ensemble. We also trained a gpt-2 model, which needed quite a bit of tweaks to get to a reasonable score. It was still only around 0.396 for a single model, but added to the ensemble nonetheless. Our code for transformers is available at <a href="https://github.com/lopuhin/kaggle-jigsaw-2019">https://github.com/lopuhin/kaggle-jigsaw-2019</a>. Overall we found <a href="https://github.com/huggingface/pytorch-transformers">https://github.com/huggingface/pytorch-transformers</a> really nice to use. It’s clear that transformers are now a must in any NLP competition, and we’re still waiting for more tricks to be discovered on how to use them.</p> <h3>TF-IDF</h3> <p>To make models even more diverse, we trained several models on two different TF-IDF matrices. To do this, we used two kernels from the previous competition: [<a href="https://www.kaggle.com/chernyshov/logistic-regression-with-preprocessing">2</a>, <a href="https://www.kaggle.com/coolcoder22/simple-logisticregression">3</a>]. For boosting, we also tuned some parameters on the validation set. Finally, for this part we got 4 models:</p> <ul> <li><a href="https://www.kaggle.com/blackitten13/fork-of-train-preprocessing-for-tf-idf">2</a> <a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">LightGBM</a> models</li> <li><a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">Logistic Regression</a> model</li> <li><a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">NB-SVM</a> model</li> </ul> <h3>Blending</h3> <p>And this is where the magic begins. At some point we noticed that we got a better quality of validation/public set, if we change the coefficients of the models in the blend - and even make them look like very strange constants and <strong>even negative</strong>. In order not to do it by hands, we wrote a script for tuning them which you can see in the validation part of our solution. The idea is to initialize weights for models, and then change them iteratively one by one each iteration - specifically, each weight is represented as a fraction, and each time we try to change its numerator up or down by some integer values. The algorithm finishes when on the last iteration there were no changes in weights.</p> <p><em><strong>What was tried but didn't work</strong></em></p> <ul> <li>scipy optimization of coefficients (gave trivial irrelevant results and worked for much time, though we could have made a mistake somewhere)</li> <li>stacking (this was a surprise - neither LR nor LightGBM could beat the blending approach)</li> <li>other types of blending (geometric, harmonic, rank)</li> <li>a mad idea ~5 hours before the deadline - including extracted 4 simple features such as <code>clean_word_len</code>, <code>clean_char_len</code>, <code>clean_chars</code>, <code>clean_chars_ratio</code> as the models in the blend. Obviously, they had AUC-ROC ~0.5, however, they gave some boost on validation 😁 There was no boost on public, though, so we decided to throw them out of harm's way.</li> </ul> <h1>The results</h1> <p>The general results are available in the following notebooks:</p> <p><a href="https://www.kaggle.com/blackitten13/lstm-cp-clip-nb-tw-tfidf-berts-gpt-val2-best-r?scriptVersionId=16349926">Validation notebook (version 2)</a> <a href="https://www.kaggle.com/blackitten13/lstm-cp-clip-nb-tw-tfidf-bert4x-gpt-test-best?scriptVersionId=16367978">Final notebook with submission (version 3)</a></p> <p>In the final solution, we had 21 models. On the table below you can find the results of each model.</p> <p>\| Model \| Validation score \| Blend coefficient \| --- \| --- \| ---\| \| <a href="https://www.kaggle.com/blackitten13/bilstm-crawl-paragram-6-ep-0-train">LSTM CP (seed 1, epoch 3)</a> \| 0.93335312 \| 0.0316 = 50/130 * 17/207 \| LSTM CP (seed 2, epoch 3) \| 0.93439617 \| 0.0253 = 40/130 * 17/207 \| LSTM CP (seed 3, epoch 3) \| 0.93436607 \| 0.0253 = 40/130 * 17/207 \| <a href="https://www.kaggle.com/blackitten13/bilstm-c-p-target-clip-0-05-5-ep-0-train-val?scriptVersionId=16115529">LSTM clip (seed 1, epoch 2)</a> \| 0.93341418 \| 0.0514 = 40/124 * 33/207 \| LSTM clip (seed 2, epoch 3) \| 0.93333687 \| 0.0514 = 40/124 * 33/207 \| LSTM clip (seed 3, epoch 3) \| 0.93372217 \| 0.0566 = 44/124 * 33/207 \| <a href="https://www.kaggle.com/blackitten13/bilstm-numberbatch-0-train-val">LSTM nb (seed 1, epoch 4)</a> \| 0.92830396 \| -0.0204 = /4040/123 * -13/207 \| LSTM nb (seed 2, epoch 4) \| 0.92741954 \| -0.0199 = 39/123 * -13/207 \| LSTM nb (seed 3, epoch 4) \| 0.92771196 \| -0.0225 = 44/123 * -13/207 \| LSTM tw (seed 1, epoch 3) \| 0.93013805 \| -0.0016 = 40/119 * -1/207 \| <a href="https://www.kaggle.com/blackitten13/bilstm-twitter-1-train-val">LSTM tw (seed 2, epoch 3)</a> \| 0.92989958 \| -0.0016 = 39/119 * -1/207 \| LSTM tw (seed 3, epoch 3) \| 0.93067808 \| -0.0016 = 40/119 * -1/207 \| <a href="https://www.kaggle.com/blackitten13/fork-of-train-preprocessing-for-tf-idf">TF-IDF LightGBM #1</a> \| 0.88133766 \| -0.0037 = 42/164 * -3/207 \| <a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">TF-IDF LightGBM #2</a> \| 0.88202771 \| -0.0036 = 41/164 * -3/207 \| <a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">TF-IDF LR</a> \| 0.88507331 \| -0.0039 = 44/164 * -3/207 \| <a href="https://www.kaggle.com/blackitten13/simple-tf-idf-models-train-val">TF-IDF NB-SVM</a> \| 0.88679439 \| -0.0033 = 37/164 * -3/207 \| BERT Uncased \| <strong>0.94027252</strong> \| 0.2986 = 59/126 * 132/207 \| BERT Cased \| 0.93841706 \| 0.1518 = 30/126 * 132/207 \| BERT Uncased Fresh \| 0.93919279 \| -0.1417 = -28/126 * 132/207 \| BERT Large \| 0.94019660 \| <strong>0.3290 = 65/126 * 132/207</strong> \| GPT-2 \| 0.93866539 \| 0.2029 = 42/207 \| <strong><em>Megablend</em></strong> \| <strong><em>0.94379378</em></strong></p> <p>To conclude: personally, this was a very useful and interesting competition for me. I learned a lot, tried many things, worked in a great team, and, last but not least, this is my first ever medal at Kaggle! (and it is not even bronze - it is silver!) Also it is a second medal for Artem, meaning he is an Expert now!🎉 🎉 </p> <h1>Bonus: an interesting bug</h1> <p>A few days before the deadline I figured out a very strange LSTM models behaviour - some of them gave different validation scores if I put them in different kernels. It was even like this: say there are models m1 m2 m3. If I put m1 in one kernel, m2 in second kernel and m3 in third, they give some results. If instead I put all them in one kernel in the following order: m1 m2 m3, then m1 has the same result, but m2 and m3 - different. Wtf? I debugged it the whole day, but then gave up and asked the teammates - what can be wrong? As Konstantin noticed, I forgot to switch my models to the inference mode with <code>model.eval()</code> and <code>with torch.no_grad():</code>. As a result Dropout and BatchNorm were in the training mode, and there produced this strange behaviour: as I understood, turned-on Dropout in m1 modified random seed somehow, causing different results in m2 and m3. Anyway, I think that from now on I will never forget to put <code>model.eval()</code> and <code>with torch.no_grad():</code> in my code😃 </p>
Jigsaw Unintended Bias in Toxicity Classification	10th place late writeup[code available now]	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Unintended Bias in Toxicity Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>The improvement from pretraining is significant in the beginning, and it was about 0.002. However, after we found more tricks, such as better loss function, the improvement was getting more and more insignificant.</p> <p>Actually, we included numeric toxic label in our auxiliary loss, and it improved our score too.</p> <p>First, many thanks to my great teammate <a href="https://www.kaggle.com/yangddd">yangDDD</a>. We entered this competition quite late, but we managed to harness the beasts.</p> <p>github url: <a href="https://github.com/thuwyh/Jigsaw-Unintended-Bias-in-Toxicity-Classification">https://github.com/thuwyh/Jigsaw-Unintended-Bias-in-Toxicity-Classification</a></p> <h2>1 Models</h2> <h3>BERT</h3> <p>In this competition, we used six BERT models. They are 1. BERT-large-uncased (further pretrained with competition data), 2. BERT-large-cased (further pretrained with competition data), 3. BERT-large-cased (trained with uncased data), 4. BERT-large-wwm-uncased (further pretrained with competition data), 5. BERT-large-wwm-cased, 6. BERT-based-uncased (further pretrained with competition data).</p> <p>To add diversity to our models, we trained BERT-large-cased model with uncased data and count it as one of our models.</p> <h3>GPT-2</h3> <p>This model didn’t achieve a high score as we expected, but it still improved the score of our ensemble model.</p> <h3>RNN</h3> <ul> <li>Glove 300d embedding</li> <li>Crawl 300d embedding</li> <li>Wordvec 100d embedding</li> <li>char-level count vector and keyword-based features</li> <li>3 RNN layers + attention pooling + classifier</li> </ul> <h2>2 Blending</h2> <p>Blend the above 6 BERT models with weights: 0.25, 0.15, 0.2, 0.2, 0.12, 0.08.</p> <p>Blend 13 RNN models with average.</p> <p>Blend the Blended BERT, RNN and GPT-2 with weights: 0.88, 0.06, 0.06.</p> <h2>3 Training method</h2> <h3>Nvidia Apex and bucket sampler</h3> <p>The typical training time of one epoch BERT-base model is about 2 hours with a single V100 GPU.</p> <h3>MTL custom loss</h3> <p>We used auxiliary labels including severe_toxicity, obscene, identity_attack, insult and threat. </p> <p>For the original toxic label, we used 0.5 as threshold to get the binary label, and then calculated the binary cross entropy with our prediction as loss1. For auxiliary labels, we directly calculated the binary cross entropy with our predictions as loss2 (both labels and logits were float numbers in the range of [0,1]). The final loss = loss1 * 0.02 + loss2. We tried a lot of other loss functions to fit the metric, and this one worked best.</p> <h3>Sentence clipping</h3> <p>If a sentence is longer than Lmax, we concatenated the 0.25Lmax tokens in the beginning and 0.75Lmax tokens in the end, for the most important information often exists in the tail and the head. The max length Lmax was set to 220.</p> <h3>Different learning rate among the layers of BERT</h3> <p>We set the learning rate of last layer of BERT encoder and the task-specific classifier to a base learning rate, such as 5e-5; and we reduced the learning rate by a ratio of 0.98 for the preceding BERT layers. As a result, the learning rate of the first layer of a BERT base model is 5e-5 * 0.98^12 . </p> <h3>Gradient accumulation</h3> <p>For BERT-base and GPT-2, we set batch size to 48 and accumulating step to 3; and for BERT-large the batch size and accumulating step were 12 and 20 respectively. </p> <p>what is correction?</p> <p>-- sorry, did not see the photo. correction is a spell checking function</p>
Jigsaw Unintended Bias in Toxicity Classification	11th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Unintended Bias in Toxicity Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>Can't remember the exact score. For bert uncased the lb score from 0.936+ to 0.943.</p> <ol> <li>yes, pregenerate the data first then finetuning.</li> <li>just jigsaw data.</li> <li>In my experiment, the loss drop fast at first. And I finetune all models (base and large) with 2 epochs.</li> </ol> <p>I use the default parameters. 😑 Actually I didn't notice the loss... Just try the finetuned lm directly.</p> <p>Thanks to kaggle and Jigsaw for this great competition and I'm almost wining my Titan RTX card😔 . After team up with @tides, we learned a lot from each other and the lb score increase a lot by simply averaging our submissions. We figured out that the diversity of our methods lead to the increase. I focus on language model finetune while tides focus on more reasonable loss function. </p> <h1>Summary</h1> <h1>toxu:</h1> <p>I didn't spent much time on custom loss function in this competition. I mainly focused on train more models with different kinds of language model and tried to make the inference faster. </p> <h2>Model</h2> <p>several rnn-based models with different embeddings(glove, fasttext, word2vec, para and their concatenation). Achieve lb:0.938+ by averaging all results.</p> <p>Bert models:</p> <ul> <li>uncased bert base</li> <li>cased bert base</li> <li>uncased bert large</li> <li>cased bert large</li> </ul> <p>All above bert models with lm finetune. I found lm finetune helps a lot but time consuming.</p> <h2>things work for me</h2> <ul> <li>preprocessing (from <a href="https://www.kaggle.com/christofhenkel/how-to-preprocessing-for-glove-part1-eda">public kernel</a>)</li> <li>lm finetune.</li> <li>pad the sentence with different length and put the sentences with similiar length in a batch. It also helps us put more models in the inference time.</li> </ul> <h2>things not work for me</h2> <ul> <li>lexvec embedding</li> <li>NumberBatch embedding</li> <li>stacking based on previous jigsaw competition</li> <li>statistical feature for sentences</li> </ul> <h1>tides:</h1> <h2>Model summary</h2> <ul> <li>bilstm + cnn: 12 with different embeddings and loss (0.94163) + 6 with differenct tokenizers</li> <li>3 bert uncased base (best 0.94346)</li> <li>1 bert cased base (0.94184)</li> <li>1 bert uncased large</li> <li>1 gpt2</li> </ul> <p>And for bert, actually when using bucketing technique, I made a mistake, which cause '[SEP]' is trimmed. After I modified the error and added another trick, the score improved locally from 0.9390 -> 0.9410 (95w train, 20w validation). But I don't have time to update models. </p> <h2>Loss:</h2> <p>I focus more on loss function. After some experiments, I try to plot distribution of predictions for 4 samples(sp, sn, bp, bn) and design loss to adjust the distributions. Following are desinged loss. Validation for each part is not so solid, so be careful.</p> <p>```python class MyFocalLoss(nn.Module): def <strong>init</strong>(self, alpha=1, gamma=1, logits=True, reduce=True): super(FocalLoss, self).<strong>init</strong>() self.alpha = alpha self.gamma = gamma self.logits = logits self.reduce = reduce</p> <pre><code>def forward(self, inputs, targets, weight, focal_weights, penalty_weights, focal_alpha, penalty_alpha): if self.logits: BCE_loss = nn.BCEWithLogitsLoss(weight=weight, reduce=False)(inputs, targets) else: BCE_loss = nn.BCELoss(weight=weight, reduce=False)(inputs, targets) pt = torch.exp(-BCE_loss) # The designed part bias_weights = (1.0 + targets.detach() - torch.sigmoid(inputs).detach()) F_loss = (1-pt)*(focal_weights) BCE_loss * (bias_weights ** penalty_weights) * focal_alpha * penalty_alpha if self.reduce: return torch.mean(F_loss) else: return F_loss </code></pre> <p><code> </code>python IDENTITY_COLUMNS_WEIGHTS= { 'black': 4.5, 'white': 3.5, 'homosexual_gay_or_lesbian': 5, 'muslim': 2.5, 'jewish': 0.2, 'female': 1, 'christian': -0.1, 'male': 0.5, 'psychiatric_or_mental_illness': 4.5 } def get_sample_weight_and_scale(train): sample_weights = np.ones((len(train),))</p> <pre><code># sn(subgroup, negtive) samples will have more weights, the weights are adjusted to balance 'bnsp_auc' and 'bpsn_auc' for i, ic in enumerate(IDENTITY_COLUMNS_WEIGHTS): sample_weights += ((train['target'].values < 0.5) & (train[ic].fillna(0) >= 0.5)).astype(int) * IDENTITY_COLUMNS_WEIGHTS[ic] # The sample receiving more annotators is more important coeff1 = np.log10(train['toxicity_annotator_count']) # The sample receiving more identity involved annotations, is more important coeff2 = np.log(train['identity_annotator_count'] + 1) / np.log(4) sample_weights = sample_weights.values * coeff1 * ((coeff2 + 1) ** 0.3) sample_weights /= sample_weights.mean() return sample_weights.values </code></pre> <p>```</p> <h2>Multiple targets:</h2> <ul> <li>MAIN_COLUMNS = ['target'], designed loss</li> <li>AUX_COLUMNS = ['target', 'insult', 'threat', 'sexual_explicit', 'obscene'], original BCE loss</li> <li>Another trick that not used in final version: AUX_COLUMNS2 = ['target1', 'target2', 'target3', 'target4', 'target5', 'target6'], original BCE loss <code>python train['target6'] = np.where(train['target'] >= 0.6, 1, 0) train['target5'] = np.where(train['target'] >= 0.5, 1, 0) train['target4'] = np.where(train['target'] >= 0.4, 1, 0) train['target3'] = np.where(train['target'] >= 0.3, 1, 0) train['target2'] = np.where(train['target'] >= 0.2, 1, 0) train['target1'] = np.where(train['target'] >= 0.1, 1, 0) </code></li> </ul> <h2>About blending:</h2> <p>The score decrease when we add more bert models. I notice the fraction of lstm based model should not be too low. So I keep it at about 0.25.</p>
Jigsaw Unintended Bias in Toxicity Classification	14th place solution	Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge	<\|system\|>You can successfully explain and teach others topics and concepts from Kaggle competition solution write-ups. You are a highly talented data scientist who has shown exceptional performance in this data-science Kaggle competition: Jigsaw Unintended Bias in Toxicity Classification <\|user\|>Explain and teach me concepts from Kaggle competition solution write-ups that solved this challenge <\|assistant\|><p>That's right.</p> <p><a href="/httpwwwfszyc">@httpwwwfszyc</a> Sorry for late reply. We have trained GBDT models individually. For example, it can be handled with the following minor modifications.</p> <p>``` class CustomBertClassifier(nn.Module): ... def forward(self, x): ...</p> <h2> return pooled_output, out</h2> <p>embeddings = np.zeros((len(X), 768))</p> <p>for i, (x_batch, ids) in enumerate(loader): x_batch = trim_tensors([x_batch])[0]</p> <pre><code>with torch.no_grad(): emb, pred = model(x_batch.to(device), attention_mask=(x_batch &gt; 0).to(device), labels=None) ... embeddings[i * batch_size: (i + 1) * batch_size] = emb.detach().cpu() </code></pre> <hr> <h1>train/predict GBDT models with <code>embeddings</code> as feature</h1> <p>... ```</p> <ol> <li>Simply remove last linear layer on each BERT model.</li> <li>Improve score about 0.0003 ~ 0.0005 on private LB.</li> </ol> <blockquote> <p>What was the LB score for the GBDT models?</p> </blockquote> <p>About 0.9437~0.9439 on public LB. Their are close to our BERT models, but the rank correlation are not very high.</p> <blockquote> <p>Were rank averages better than simple averages?</p> </blockquote> <p>We tried only rank averaging. Because we checked good LB score using rank averaging at an early stage.</p> <p>First of all, We would like to thank the competition hosts for such a great competition. Also, congratulate all winers and participants.</p> <h3>Models</h3> <p>Our solution is the weighted rank averaging of the following models. We didn't try XLNet because there was not enough time for us...</p> <ul> <li>RNN / lstm-gru-Attention</li> <li>BERT-Base uncased</li> <li>BERT-Large uncased</li> <li>GPT-2</li> <li>LightGBM w/ BERT-Base uncased 768d feature</li> <li>LightGBM w/ BERT-Large uncased 1024d feature</li> <li>CatBoost w/ BERT-Base uncased 768d feature</li> <li>CatBoost w/ BERT-Large uncased 1024d feature</li> </ul> <h3>Preprocessing</h3> <ul> <li><p>RNN</p> <ul><li>spell correct to match more words to embedding</li> <li>do not apply contraction mapping</li></ul></li> <li><p>BERT/GPT-2</p> <ul><li>apply spell correction and contraction mapping</li></ul></li> </ul> <h3>Embedding</h3> <ul> <li>RNN <ul><li>fastText & train fastText (300d, on the competiton dataset)</li> <li>Glove & train fastText</li></ul></li> </ul> <h3>Loss function</h3> <p>We used weighted FocalLoss and <a href="https://www.kaggle.com/matsuik/subgroup-negative-weighting">subgroup negative weighting</a> for all NN models.</p> <h3>What didn't work for us</h3> <ul> <li>data augmentation (translation, past competition data)</li> <li>more spell correction for toxic words about 2,000 patterns (We didn't use most.)</li> <li>oof for indicator feature (completion, residual)</li> <li>trained RNN with part of sentences (first/last 50 characters)</li> <li>use statistical features for BERT, GPT-2 and GBDT</li> </ul> <p>That's correct.</p> <p>Just return <code>pooled_output</code> and use them to train and predict. There are not any special processing.</p>