README.md

metadata

license: mit
tags:
  - code
size_categories:
  - n<1K

Gists

This 🤗 dataset contains some of my GitHub Gists at https://gist.github.com/alvarobartt, ported here so that its cleaner and easier to maintain.

Available gists

• causallm-to-hub.py: to upload any AutoModelForCausalLM to the 🤗 Hub from a local path, useful after some LLM fine-tuning, as sometimes accelerate gets stuck while pushing to the Hub, so I tend to do that in a separate process after each epoch has been dumped into the disk.

• dpo-qlora-4bit.py: to fine-tune an AutoModelForCausalLM using Q-LoRA in 4-bit, in this case the fine-tuning is done using 🤗 trl.DPOTrainer built on top of transformers useful for intent alignment of LMs on low resources, ~80GB of VRAM.