Google QUEST Q&A Labeling 1st place solution

Below you can find an outline of how to reproduce our solution for the Google QUEST Q&A Labeling competition. If you run into any trouble with the setup/code or have any questions please contact me at [email protected].

The solution is also described in this post on Kaggle, the inference part is fully reproduced in this Kaggle Notebook.

Archive contents

The archive contains the following files:

• stackx-base-cased.tar.gz – pretrained BERT language model, output of step 1 in the "Model training" section
• sampled_sx_so.csv.gz – chunk of StackExchange dump used to generate pseudo-labels
• pseudo-100k-3x-blend-no-leak.tar.gz – pseudo-labels, output of step 2 in the "Model training" section

Hardware

• 1 x NVIDIA Quadro P6000
• 2 x NVIDIA 1080 Ti
• 5 x NVIDIA 1080 Ti (only for language model training)

Software and environments

• Conda 4.7.10 with Python 3.6.6
• CUDA 10.0.130
• cuDNN 7.5.0
• NVIDIA drivers v. 418.67

We run all experiments in a Conda environment and provide a full list of required packages requirements_full.txt. To do the same, run:

• conda create -n qa_quest_env python=3.6.6
• conda activate qa_quest_env
• pip install -r requirements_full.txt

However, there are a lot of packages listed there which are not much relevant to this project. A minimal list of requirements is specified in requirements_minimal.txt which was formed with the pipreqs utility. This handy utility crawls project source code and lists all required packages. So you can run pip install -r requirements_minimal.txt and then install missing packages on the go (if any).

Apart from pip-installable packages, we use a custom lightweight library called mag to keep track of experiments. Also our installation of the fairseq library is a bit different (actually, it's hacked for our needs). These two can be installed by running sh bash/setup.sh

Model training

For some of our models, we perform language model finetuning with StackExchange data. Then we run 5-fold cross-validation for 4 models (2 BERT ones, one RoBERTa, and one BART) averaging predictions of all 5 model checkpoints for each model type. Finally, blending 4 predictions. In this section, we cover everything related to model training:

1. Language model finetuning with StackExchange data
2. Generating pseudo-labels
3. Training BERT-base-cased pretrained with StackExchange
4. Training BERT-base-cased pretrained with StackExchange + pseudo-labels
5. Training RoBERTa-base with pseudo-labels
6. Training BART-large with pseudo-labels

1. Language model finetuning with StackExchange data

1a. Scraping and processing StackExchange questions and answers

For this purpose, we download and process StackExchange dumps, to reproduce these steps from scratch for a small subsample, run sh bash/training/train1a_prepare_stackx_data.sh. It the same as running the following two commands:

• python step1_lm_finetuning/data_preparation/scrape_stack_exchange.py
• python step1_lm_finetuning/data_preparation/clean_stack_exchange_qa.py

See comments in these files. Also, you can find 2 corresponding Jupyter notebooks in the same folder, just for demonstration of the scraping/cleaning process.

Output is written to the input/qa_stackexchange_cleaned.csv file. This files is needed for the next step.

1b. Fine-tuning BERT language model with StackExchange data:

sh bash/training/train1b_train_bert_stackx_lang_model.sh – this runs BERT language model fine-tuning with StackExchange data from the previous step (input/qa_stackexchange_cleaned.csv).

This script writes model checkpoints and training logs to input/stackx-base-cased, the following two BERT models use this checkpoint. The checkpoint is also shared as stackx-base-cased.tar.gz in the archive. To go on reproducing results, it's better to unzip contents of stackx-base-cased.tar.gz into input/stackx-base-cased.

2. Generating pseudo-labels

sh bash/pseudo/create_all_pseudo_labels_toy.sh – this runs 3 basic models (bert-base, bert-large, and bert-base-pretrained) on the competition data (a toy example, first 50 rows), then creates pseudo-labels using these 3 models, the result is stored in the pseudo-predictions/pseudo-100k-3x-blend-no-leak/ folder. Run sh bash/pseudo/create_all_pseudo_labels_toy.sh (without _toy) for the actual generation of pseudo-labels. This requires sampled_sx_so.csv.gz from the shared archive and results in pseudo-100k-3x-blend-no-leak.tar.gz shared in the same archive.

3. BERT-base-cased pretrained with StackExchange

Training 5 BERT-base models (cross-validation): sh bash/training/train3_bert_base_cased_stackx_pretrained.sh toy. The result (one model checkpoint for each fold) is written to input/model1_ckpt. Full training is done without the toy argument, the result is found in this Kaggle Dataset.

4. BERT-base-cased pretrained with StackExchange + pseudo-labels

Training 5 BERT-base models (cross-validation): sh bash/training/train4_bert_base_cased_stackx_with_pseudo_labels.sh toy. The result is written to experiments/1-8-5-head_tail-pseudonoleakrandom100k-1e-05-210-260-500-26-200. Full training is done without the toy argument, the result is found in this Kaggle Dataset.

5. RoBERTa-base with pseudo-labels

Download RoBERTa-base checkpoint by running sh bash/training/load_roberta_weights.sh

Training 5 RoBERTa-base models (cross-validation): sh bash/training/train5_roberta_with_pseudo_labels.sh. The result is written to experiments/2-4-5-head_tail-pseudonoleakrandom100k-1e-05-210-260-500-26-roberta-200 and can also be found in this Kaggle Dataset Also, here 5 model checkpoints (one per each fold) are stored.

6. BART-large with pseudo-labels

Training 5 BART-large models (cross-validation): sh bash/training/train6_bart_with_pseudo_labels.sh toy. The result is written to experiments/4-2-5-head_tail-bart-2e-05-210-260-500-26-split_pseudo-250. Full training is done without the toy argument, the result is found in this Kaggle Dataset, and here 5 model checkpoints (one per each fold) are stored.

Inference

These are the steps to reproduce our final solution (same as our Kaggle Notebook 1st place solution). You can just run sh bash/inference/run_inference.sh (execution log is found in logs/all_inference.log, takes ~10 minutes) to run the whole pipeline (from data loading to forming a submission file). Here is a breakdown of all steps:

1. Make sure you've got a fresh Kaggle API token and download competition data by running sh bash/download_comp_data.sh. This will populate input/google-quest-challenge with three more CSV files - train.csv, test.csv and sample_submission.csv
2. Run sh bash/download_all_model_ckpts_for_inference.sh. This will download all models needed for inference (about 18 Gb, might take from several minutes to more that an hour depending on Internet speed):

• BERT checkpoints from this Dataset (the result of running steps 1, 3 above)
• BERT checkpoints from this Dataset (the result of running steps 2, 4 above)
• RoBERTa checkpoints from this Dataset (the result of running steps 1, 2, 5 above)
• BART checkpoints from this Dataset (the result of running steps 2, 6 above)

2. Inference with 5 checkpoints of BERT-base-cased finetuned with StackExchange data: sh bash/inference/model1_inference.sh
3. Same for the BERT model with pseudo-labels: sh bash/inference/model2_inference.sh
4. Inference with with 5 checkpoints of RoBERTa finetuned with StackExchange data, with pseudo-labels: sh bash/inference/model3_inference.sh
5. Inference with with 5 checkpoints of BART with pseudo-labels: sh bash/inference/model4_inference.sh
6. Once inference is done, final steps include blending, and postprocessing model predictions: sh bash/blending_n_postprocessing.sh

Final submission submissions/submission.csv scores 0.46893 on the public competition leaderboard. To actually submit a CSV file, you can check this Kaggle Notebook.