Skip to content

Latest commit

 

History

History
13 lines (12 loc) · 1.01 KB

README.md

File metadata and controls

13 lines (12 loc) · 1.01 KB
Raw

Natural Language Unerstanding on GLUE

This directory contains the scripts to finetune RoBERTa-Base and RoBERTa-Large on the GLUE benchmark: General Language Understanding Evaluation. Based on the script run_glue_no_trainer.py.

The following scripts will run the hyperparameter search (with seed 42) and final run with 5 seeds (42, 43, 44, 45, 46). GLUE dataset will downloaded automatically using HuggingFace datasets library.

# RoBERTa-Base
./run_glue_base.sh <TASK_NAME> <GPU_ID> <EPOCHS> --gift_rank 32
# RoBERTa-Large
./run_glue_large.sh <TASK_NAME> <GPU_ID> <EPOCHS> --gift_rank 32
  • <TASK_NAME> can take values cola, mrpc, qnli, rte, sst2 or stsb
  • The number of training epochs for each task is given in Table 9 in the Appendix.
  • <GPU_ID> is the numeric ID of the GPU to be used (e.g., for cuda:0, use <GPU_ID> as 0)