fine-tuning-large-language-models-on-limited-hardware

NYU Tandon, ECE-GY 9143: High Performance Machine Learning, End Semester Project

Project:

Optimize the process of domain adaptation in natural language processing, i.e., fine-tuning large language models for a particular domain on limited hardware . This is done by using 8-bit quantization, LoRA and other techniques.

Repository:

• Fine_Tuning.ipynb: Main notebook for training and evaluation
• train_v4.py: Same script as ipynb in python script form, for submitting a sbatch job
• Inference.ipynb: Inference notebook, for generating text from a saved model
• shell scripts/data_download.sh: Script for downloading data
• shell scripts/run_train_job.sh: Script for running the train job (automatically runs train_v4.py)
• gpt2_logs: Training and validation logs for the GPT-2 fine tuning run
• opt_logs: Training and validation logs for the OPT fine tuning run
• sbatch_job_log: Logs from the sbatch job for GPT-2 fine tuning run

How to run the code:

• Create sbatch job and run run_train_job.sh (16 cores, 60 GB RAM, 1 RTX GPU)

Results:

Achieved a perplexity score of 7.24 while fine-tuning GPT-2 model (1.5B param model) on the NiH grants dataset after 17 epochs. Qualitatively, the text generated is plausibly similar to an NIH grant.