This repository contains a PyTorch implementation of the electra model from the paper

ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

by Kevin Clark. Minh-Thang Luong. Quoc V. Le. Christopher D. Manning

NOTE： 🤗This version is experience version,and the offical PyTorch version is waiting for the update of 🤗huggingface

NOTE: 2020-04-08 ELECTRA is now available in PyTorch through the HuggingFace Transformers library! https://github.com/huggingface/transformers/releases/tag/v2.8.0

• pytorch=1.10+
• cuda=9.0
• cudnn=7.5
• scikit-learn
• sentencepiece
• python3.6+

English: Official download links: google electra

Chinese:

• https://github.com/CLUEbenchmark/ELECTRA
• https://github.com/ymcui/Chinese-ELECTRA

１. Place config.json into the prev_trained_model/electra_base directory. example:

├── prev_trained_model
|  └── electra_base
|  |  └── pytorch_model.bin
|  |  └── config.json
|  |  └── vocab.txt

2．convert electra tf checkpoint to pytorch

python convert_electra_tf_checkpoint_to_pytorch.py \
    --tf_checkpoint_path=./prev_trained_model/electra_large \
    --electra_config_file=./prev_trained_model/electra_large/config.json \
    --pytorch_dump_path=./prev_trained_model/electra_large/pytorch_model.bin

Before running anyone of these GLUE/CLUE tasks you should download the GLUE data /CLUE data by running script named download_xxxx_data in the directorytools and unpack it to some directory $DATA_DIR.

3．run sh scripts/run_classifier_sst2.shto fine tuning albert model

Performance of electra on GLUE benchmark results using a single-model setup on dev:

Performance of electra on CLUE benchmark results using a single-model setup on dev:

Performance of electra on chnsenticorp results using a single-model setup on dev:

test on small dataset.