|
--- |
|
language: en |
|
pipeline_tag: fill-mask |
|
license: cc-by-sa-4.0 |
|
tags: |
|
- legal |
|
model-index: |
|
- name: lexlms/roberta-base |
|
results: [] |
|
widget: |
|
- text: "The applicant submitted that her husband was subjected to treatment amounting to [MASK] whilst in the custody of police." |
|
datasets: |
|
- lexlms/lex_files |
|
--- |
|
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You |
|
should probably proofread and complete it, then remove this comment. --> |
|
|
|
# LexLM base |
|
|
|
This model was continued pre-trained from RoBERTa base (https://huggingface.co/roberta-base) on the LeXFiles corpus (https://huggingface.co/datasets/lexlms/lexfiles). |
|
|
|
## Model description |
|
|
|
LexLM (Base/Large) are our newly released RoBERTa models. We follow a series of best-practices in language model development: |
|
* 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 |
|
|
|
More information needed |
|
|
|
## Training and evaluation data |
|
|
|
The model was trained on the LeXFiles corpus (https://huggingface.co/datasets/lexlms/lexfiles). For evaluation results, please consider our work "LeXFiles and LegalLAMA: Facilitating English Multinational Legal Language Model Development" (Chalkidis* et al, 2023). |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 0.0001 |
|
- train_batch_size: 32 |
|
- eval_batch_size: 32 |
|
- seed: 42 |
|
- distributed_type: tpu |
|
- num_devices: 8 |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 512 |
|
- total_eval_batch_size: 256 |
|
- 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 |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | |
|
|:-------------:|:-----:|:-------:|:---------------:| |
|
| 1.0389 | 0.05 | 50000 | 0.9802 | |
|
| 0.9685 | 0.1 | 100000 | 0.9021 | |
|
| 0.9337 | 0.15 | 150000 | 0.8752 | |
|
| 0.9106 | 0.2 | 200000 | 0.8558 | |
|
| 0.8981 | 0.25 | 250000 | 0.8512 | |
|
| 0.8813 | 1.03 | 300000 | 0.8203 | |
|
| 0.8899 | 1.08 | 350000 | 0.8286 | |
|
| 0.8581 | 1.13 | 400000 | 0.8148 | |
|
| 0.856 | 1.18 | 450000 | 0.8141 | |
|
| 0.8527 | 1.23 | 500000 | 0.8034 | |
|
| 0.8345 | 2.02 | 550000 | 0.7763 | |
|
| 0.8342 | 2.07 | 600000 | 0.7862 | |
|
| 0.8147 | 2.12 | 650000 | 0.7842 | |
|
| 0.8369 | 2.17 | 700000 | 0.7766 | |
|
| 0.814 | 2.22 | 750000 | 0.7737 | |
|
| 0.8046 | 2.27 | 800000 | 0.7692 | |
|
| 0.7941 | 3.05 | 850000 | 0.7538 | |
|
| 0.7956 | 3.1 | 900000 | 0.7562 | |
|
| 0.8068 | 3.15 | 950000 | 0.7512 | |
|
| 0.8066 | 3.2 | 1000000 | 0.7516 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.20.0 |
|
- Pytorch 1.12.0+cu102 |
|
- Datasets 2.6.1 |
|
- Tokenizers 0.12.0 |
|
|