- Notifications
You must be signed in to change notification settings - Fork5
Kel-Lu/SciGen
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
SciGen is a generation model trained on scientific articles based on GPT2 and the code is based heavily on HuggingFace's GPT2 transformers examples. For more information see our paperExplaining Relationships Between Scientific Documents
We note thatSciGPT2_Clean was trained on a reduced set of papers to prevent leakage in our experiments and is released for reproducibility. In general, we recommend using the full veresion ofSciGPT2.
Please follow the steps underdata processing.
python ft.py --model_type=gpt2 --do_eval --max_eval_steps 100000 --num_train_epochs=1 --save_steps=5000 --eval_all_checkpoints --tokenizer_path=$MODEL_PATH --output_dir=$OUTPUT_PATH --eval_data_file=$EVAL_FILE --model_name_or_path=$MODEL_PATH
python val_generation.py --model_type=gpt2 --length 50 --stop_token='. ' --tokenizer_path=$TOKENPATH --prompt=$TEST_FILE --output_file $OUTPUT_FILE --model_name_or_path=$MODEL_PATH