README.md

Language Models Can Learn Exceptions to Syntactic Rules (Leong and Linzen 2023)

This repository contains code for the paper Language Models Can Learn Exceptions to Syntactic Rules.

Repository Contents

• 100M words training dataset based on OpenWebText (Gokaslan and Cohen 2019): data/100M
• Final checkpoints of 5 models trained with early stopping for 30 epochs: models/100_[n]
• Training script: train.py
• Test sentences: data/test_sentences.csv
• Evaluation script: eval.py
• Sentence scores: scores

Instructions to replicate

Setup

conda env create -f environment.yml
conda activate traingpt

Training

To train your own model on the 100M word dataset, use the following code:

python train.py \
    --output_dir models/ \
    --model_type gpt2 \
    --tokenizer_name gpt2 \
    --train_file data/100M/train_100M.txt \
    --validation_file data/100M/validation.txt \
    --dataset_config_name unshuffled_deduplicated_no \
    --bf16 true \
    --do_train true \
    --do_eval true \
    --block_size 512 \
    --per_device_train_batch_size 16 \
    --per_device_eval_batch_size 16 \
    --gradient_accumulation_steps 5 \
    --learning_rate 6e-4 \
    --warmup_steps 100 \
    --adam_beta1 0.9 \
    --adam_beta2 0.98 \
    --weight_decay 0.01 \
    --num_train_epochs 30 \
    --logging_steps 50 \
    --save_steps 1200 \
    --evaluation_strategy "steps" \
    --eval_steps 400 \
    --load_best_model_at_end true

Evaluation

Run eval.py on a test dataset. The script looks into the models directory for a model with the given name and checkpoint and saves sentence scores in scores.

python eval.py data/test_sentences.csv --model 100M_4 --checkpoint 168000