0

unaligned

Research scripts and notes for preference optimization experiments (DPO/LoRA) on small chat models; training, eval, and inference utilities.

Overview

Unaligned is a scratchpad repo for running small‑scale Direct Preference Optimization (DPO) experiments with TinyLlama variants: training, evaluation, and simple inference/chat CLIs.

What’s inside

  • DPO training script with LoRA adapters (TinyLlama base)
  • Logging of preference metrics (accuracy, reward margins) and diagnostics
  • CLI tools for single‑prompt generation and interactive chat
  • Checkpoint management and merged‑weights export

Notes

  • CPU‑friendly tips (shorter max lengths, smaller evals)
  • Safety disclaimer: datasets may include toxic content; use for research only

Repo

GitHub: https://github.com/J0YY/unaligned