README.md

SelfContrast

Code for the paper Mitigating the Learning Bias towards Repetition by Self-Contrastive Training for Open-Ended Generation (ACL 2023 Short Findings paper)

Prerequisites

The code is written in PyTorch library. Main dependencies are as follows:

• Python: 3.6.9
• torch: 1.8.1
• transformers: 4.6.1

Other dependencies can be found in requirements.txt

Computing infrastructure

We train HINT based on the platform:

• OS: Ubuntu 16.04.3 LTS (GNU/Linux 4.4.0-98-generic x86_64)
• CUDA Version: 10.1
• GPU: NVIDIA Tesla V100

Quick Start

1. Datasets

The full data can be downloaded from THUcloud.

2. Training SelfCont

The initial checkpoint of GPT2 can be downloaded from HuggingFace. We provide our checkpoints on THUcloud.

• The 1st stage (get the premature checkpoint): Execute the following command (or run bash ./run0.sh directly):

  data_name=wikitext
  env CUDA_VISIBLE_DEVICES=0 python3 -u ./run_clm.py \
    --model_name_or_path gpt2 \
    --train_file ./${data_name}_data/train.txt \
    --validation_file ./${data_name}_data/val.txt \
    --do_train \
    --do_eval \
    --num_train_epochs 100 \
    --max_eval_samples 1000 \
    --dataloader_num_workers 64 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --gradient_accumulation_steps 2 \
    --output_dir ./${data_name}_f0_ckpt \
    --logging_steps 5 \
    --learning_rate 1e-4 \
    --lr_scheduler_type linear \
    --evaluation_strategy epoch \
    --save_strategy epoch \
    --cache ../cache

  The 1st training stage is exactly the same as fine-tuning the standard GPT2 model.

• The 2nd stage (get the final model): Execute the following command (or run bash ./run.sh directly):

  data_name=wikitext
  env CUDA_VISIBLE_DEVICES=0 python3 -u ./run_clm.py \
    --model_name_or_path ./${data_name}_f0_ckpt \
    --model_name_or_path2 ./${data_name}_f0_ckpt \
    --train_file ./${data_name}_data/train.txt \
    --validation_file ./${data_name}_data/val.txt \
    --do_train \
    --do_eval \
    --num_train_epochs 100 \
    --max_eval_samples 1000 \
    --dataloader_num_workers 64 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --gradient_accumulation_steps 2 \
    --output_dir ./${data_name}_selfcont_ckpt \
    --logging_steps 5 \
    --learning_rate 1e-4 \
    --lr_scheduler_type linear \
    --evaluation_strategy epoch \
    --save_strategy epoch \
    --cache ../cache

3. Inference

Execute the following command to generate texts (or run bash ./gen.sh directly):

data_name=wikitext
model_ckpt_path=./${data_name}_selfcont_ckpt
result_file=./result.txt
device=cuda:0
bsz=16	# batch size
task_name=test
data_ipt_file=./${data_name}_data/${task_name}.txt
topp=0	# p of top-p sampling, 0 means greedy decoding
python3 ./gen.py $model_ckpt_path $result_file $device $bsz $task_name $data_ipt_file $topp

4. Evaluation

Execute the following command for evaluation:

cd ./eval
python3 ./eval.py

You can change result_list in the script eval.py to specify the results you want to evaluate.

Citation

Please kindly cite our paper if this paper and it is helpful.

@misc{guan2023mitigating,
      title={Mitigating the Learning Bias towards Repetition by Self-Contrastive Training for Open-Ended Generation}, 
      author={Jian Guan and Minlie Huang},
      year={2023},
      eprint={2307.01542},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}