GitHub - petermchale/fine-tune-llm: Fine-tune open-source LLM on an instruction dataset

Parameter-efficient fine-tuning an open-source 7B-parameter LLM for instruction following

I fine-tuned a 7B LLM on a public instruction dataset using LoRA for 0.5 epochs over 5 hours on a single-GPU g4dn.xlarge EC2 instance.

Due to the size of the LLM, I had to increase the volume size of the EC2 instance:

(pytorch) du -hs .cache/huggingface/hub/models--Salesforce--xgen-7b-8k-base
26G     .cache/huggingface/hub/models--Salesforce--xgen-7b-8k-base

The training code uses the torch, transformers, peft and trl libraries. See requirements.txt.

The training run is tracked here and the fine-tuned model adaptors are available here.

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
.gitignore		.gitignore
README.md		README.md
infer-and-evaluate.ipynb		infer-and-evaluate.ipynb
requirements.txt		requirements.txt
run.sh		run.sh
stderr.log		stderr.log
stdout.log		stdout.log
train.py		train.py