Multilingual Legal Language Models

Legal XLM Models trained on the Multilingual Legal Pile

Monolingual base size models are available for the following languages: Bulgarian, Czech, Danish, German, Greek, English, Spanish, Estonian, Finnish, French, Irish, Croatian, Hungarian, Italian, Lithuanian, Latvian, Maltese, Dutch, Polish, Portuguese, Romanian, Slovak, Slovenian, Swedish

Monolingual large size models are available for the following languages: English, German, French, Italian, Spanish, Portuguese

Benchmarking on LEXTREME

See https://github.com/JoelNiklaus/LEXTREME

Benchmarking on LexGLUE

See https://arxiv.org/abs/2306.02069.

Code Base

Train XLM

Create a TPU VM instance with the following script:

gcloud compute tpus tpu-vm create tpu1 --zone=europe-west4-a --accelerator-type=v3-8 --version=tpu-vm-pt-1.12

Connect to the instance:

gcloud compute tpus tpu-vm ssh tpu1 --zone europe-west4-a

Set up the environment:

git clone https://github.com/JoelNiklaus/MultilingualLegalLMPretraining
cd MultilingualLegalLMPretraining
sudo bash setup_tpu_machine.sh

Put your huggingface token in data/__init__.py and in scripts/train_mlm_tpu.sh.

Make sure that you delete the output_dir locally and the huggingface model repo (hub_model_id) before training.

For TPU acceleration use the following script:

sudo sh train_mlm_tpu.sh

For GPU acceleration use the following script:

sh train_mlm_gpu.sh

Evaluate XLM

sh eval_mlm_gpu.sh

Modify pre-trained XLM-R

export PYTHONPATH=.
python3 src/mod_teacher_model.py --teacher_model_path xlm-roberta-base --student_model_path data/plms/legal-xlm-base

Longformerize pre-trained RoBERTa LM

export PYTHONPATH=.
python3 src/longformerize_model.py --roberta_model_path data/plms/legal-xlm-base --max_length 4096 --attention_window 128

Pipeline

1. Train tokenizer (Only RoBERTa needed because we convert BERT models to RoBERTa)

    export PYTHONPATH=. && python3 src/pretraining/train_tokenizer.py | tee train_tokenizer.log

2. Evaluate tokenizer

    export PYTHONPATH=. && python3 src/pretraining/evaluate_tokenizer.py | tee evaluate_tokenizer.log

3. Mod Teacher Model

    export PYTHONPATH=. NAME=legal-xlm-roberta-base SIZE=128k  && python3 src/modding/mod_teacher_model.py --teacher_model_path xlm-roberta-base --student_model_path data/plms/${NAME}_${SIZE} --output_dir data/plms/${NAME} | tee mod_teacher_model.log

4. Train MLM (monolingual: 500K steps) (TPUs or GPUs)

    sudo sh scripts/train_mlm_tpu.sh | tee train_mlm_tpu.log

or

    sh scripts/train_mlm_gpu.sh | tee train_mlm_gpu.log

5. Evaluate MLM

    sh scripts/eval_mlm_gpu.sh | tee eval_mlm_gpu.log

6. Longformerize MLM

    export PYTHONPATH=. && python3 src/modding/longformerize_model.py --roberta_model_path joelito/legal-xlm-roberta-base --longformer_model_path data/plms/legal-xlm-longformer-base | tee longformerize_model.log

7. Train Longformer MLM (monolingual: 50K steps) (only GPUs!)

    sh scripts/train_mlm_longformer.sh | tee train_mlm_longformer.log

8. Evaluate Longformer MLM

    sh scripts/eval_mlm_gpu.sh | tee eval_mlm_gpu.log

Troubleshooting

If you get a PermissionError: [Errno 13] Permission denied: set the permissions to 777.

If you get strange git lfs errors, delete the huggingface model repo and the output directory