LexLM large

Model description

• We warm-start (initialize) our models from the original RoBERTa checkpoints (base or large) of Liu et al. (2019).
• We train a new tokenizer of 50k BPEs, but we reuse the original embeddings for all lexically overlapping tokens (Pfeiffer et al., 2021).
• We continue pre-training our models on the diverse LeXFiles corpus for additional 1M steps with batches of 512 samples, and a 20/30% masking rate (Wettig et al., 2022), for base/large models, respectively.
• We use a sentence sampler with exponential smoothing of the sub-corpora sampling rate following Conneau et al. (2019) since there is a disparate proportion of tokens across sub-corpora and we aim to preserve per-corpus capacity (avoid overfitting).
• We consider mixed cased models, similar to all recently developed large PLMs.

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: tpu
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• training_steps: 1000000

Training results

Framework versions

• Transformers 4.20.0
• Pytorch 1.12.0+cu102
• Datasets 2.7.0
• Tokenizers 0.12.0

Citation

@inproceedings{chalkidis-garneau-etal-2023-lexlms,
    title = {{LeXFiles and LegalLAMA: Facilitating English Multinational Legal Language Model Development}},
    author = "Chalkidis*, Ilias and 
              Garneau*, Nicolas and
              Goanta, Catalina and 
              Katz, Daniel Martin and 
              Søgaard, Anders",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics",
    month = july,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/2305.07507",
}