MiniLM: Small and Fast Pre-trained Models for Language Understanding and Generation
MiniLM is a distilled model from the paper "MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers".
Please find the information about preprocessing, training and full details of the MiniLM in the original MiniLM repository.
Please note: This checkpoint uses BertModel
with XLMRobertaTokenizer
so AutoTokenizer
won't work with this checkpoint!
Multilingual Pretrained Model
- Multilingual-MiniLMv1-L12-H384: 12-layer, 384-hidden, 12-heads, 21M Transformer parameters, 96M embedding parameters
Multilingual MiniLM uses the same tokenizer as XLM-R. But the Transformer architecture of our model is the same as BERT. We provide the fine-tuning code on XNLI based on huggingface/transformers. Please replace run_xnli.py
in transformers with ours to fine-tune multilingual MiniLM.
We evaluate the multilingual MiniLM on cross-lingual natural language inference benchmark (XNLI) and cross-lingual question answering benchmark (MLQA).
Cross-Lingual Natural Language Inference - XNLI
We evaluate our model on cross-lingual transfer from English to other languages. Following Conneau et al. (2019), we select the best single model on the joint dev set of all the languages.
Model | #Layers | #Hidden | #Transformer Parameters | Average | en | fr | es | de | el | bg | ru | tr | ar | vi | th | zh | hi | sw | ur |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
mBERT | 12 | 768 | 85M | 66.3 | 82.1 | 73.8 | 74.3 | 71.1 | 66.4 | 68.9 | 69.0 | 61.6 | 64.9 | 69.5 | 55.8 | 69.3 | 60.0 | 50.4 | 58.0 |
XLM-100 | 16 | 1280 | 315M | 70.7 | 83.2 | 76.7 | 77.7 | 74.0 | 72.7 | 74.1 | 72.7 | 68.7 | 68.6 | 72.9 | 68.9 | 72.5 | 65.6 | 58.2 | 62.4 |
XLM-R Base | 12 | 768 | 85M | 74.5 | 84.6 | 78.4 | 78.9 | 76.8 | 75.9 | 77.3 | 75.4 | 73.2 | 71.5 | 75.4 | 72.5 | 74.9 | 71.1 | 65.2 | 66.5 |
mMiniLM-L12xH384 | 12 | 384 | 21M | 71.1 | 81.5 | 74.8 | 75.7 | 72.9 | 73.0 | 74.5 | 71.3 | 69.7 | 68.8 | 72.1 | 67.8 | 70.0 | 66.2 | 63.3 | 64.2 |
This example code fine-tunes 12-layer multilingual MiniLM on XNLI.
# run fine-tuning on XNLI
DATA_DIR=/{path_of_data}/
OUTPUT_DIR=/{path_of_fine-tuned_model}/
MODEL_PATH=/{path_of_pre-trained_model}/
python ./examples/run_xnli.py --model_type minilm \
--output_dir ${OUTPUT_DIR} --data_dir ${DATA_DIR} \
--model_name_or_path microsoft/Multilingual-MiniLM-L12-H384 \
--tokenizer_name xlm-roberta-base \
--config_name ${MODEL_PATH}/multilingual-minilm-l12-h384-config.json \
--do_train \
--do_eval \
--max_seq_length 128 \
--per_gpu_train_batch_size 128 \
--learning_rate 5e-5 \
--num_train_epochs 5 \
--per_gpu_eval_batch_size 32 \
--weight_decay 0.001 \
--warmup_steps 500 \
--save_steps 1500 \
--logging_steps 1500 \
--eval_all_checkpoints \
--language en \
--fp16 \
--fp16_opt_level O2
Cross-Lingual Question Answering - MLQA
Following Lewis et al. (2019b), we adopt SQuAD 1.1 as training data and use MLQA English development data for early stopping.
Model F1 Score | #Layers | #Hidden | #Transformer Parameters | Average | en | es | de | ar | hi | vi | zh |
---|---|---|---|---|---|---|---|---|---|---|---|
mBERT | 12 | 768 | 85M | 57.7 | 77.7 | 64.3 | 57.9 | 45.7 | 43.8 | 57.1 | 57.5 |
XLM-15 | 12 | 1024 | 151M | 61.6 | 74.9 | 68.0 | 62.2 | 54.8 | 48.8 | 61.4 | 61.1 |
XLM-R Base (Reported) | 12 | 768 | 85M | 62.9 | 77.8 | 67.2 | 60.8 | 53.0 | 57.9 | 63.1 | 60.2 |
XLM-R Base (Our fine-tuned) | 12 | 768 | 85M | 64.9 | 80.3 | 67.0 | 62.7 | 55.0 | 60.4 | 66.5 | 62.3 |
mMiniLM-L12xH384 | 12 | 384 | 21M | 63.2 | 79.4 | 66.1 | 61.2 | 54.9 | 58.5 | 63.1 | 59.0 |
Citation
If you find MiniLM useful in your research, please cite the following paper:
@misc{wang2020minilm,
title={MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers},
author={Wenhui Wang and Furu Wei and Li Dong and Hangbo Bao and Nan Yang and Ming Zhou},
year={2020},
eprint={2002.10957},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
- Downloads last month
- 14,969