pr-poet-lora

PR Poet LoRA

A LoRA adapter for Qwen/Qwen3-0.6B that turns Git commit messages into four-line rhyming poems.

It powers PR Poet, a GitHub Action that posts a poem when a pull request is opened.

This repository contains the PEFT adapter and tokenizer files, not the base model. Load it with the Qwen base model:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3-0.6B"
adapter = "mkly/pr-poet-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter)

The model was trained with Axolotl and TRL, followed by GRPO optimization for formatting, rhyme, non-repetition, commit relevance, and style. It expects the structured PR Poet chat prompt used during training rather than an arbitrary plain-text prompt.

For llama.cpp deployment, use the merged and quantized PR Poet Q4_K_M GGUF.