playbooks-lm
Subproject of Playbooks: fine-tune a small open-weight model (Qwen-3-4B / Gemma) to compile and interpret Playbooks programs locally.
A subproject of Playbooks: an end-to-end pipeline that fine-tunes a small open-weight model on real Playbooks compilation and interpretation traces, then ships it as a locally-runnable Ollama model so users can run Playbooks programs locally without frontier-model API calls. Both Qwen-3-4B and Gemma training paths exist.
The full vertical is in the repo: extract traces from LangFuse, curate prompt/response pairs with dedup, quality scoring, ChatML formatting, and stratified splits; LoRA-fine-tune on AWS SageMaker (ml.g5.xlarge, single GPU) via Docker; convert the adapter to GGUF; serve through Ollama. Configured so Playbooks can point at the local model for the hot interpreter path while keeping a frontier model for the cold compilation path. Experimental, paused alongside the main Playbooks project.