local-bread-mlx

Apple-Silicon-native Bread bake experiments for the Mac Studio.

This repo is the new local successor path for Bread/BReaDv3 ideas. It treats /Users/brixton/repos/Bread_2025 as reference architecture, not as code to rerun. The first goal is deliberately small: create a real local LoRA adapter with mlx_lm.lora, track the recipe and run metadata in git, and keep large artifacts in ignored runs/ directories until an LFS/R2 policy is verified.

Current Target

• Base model: /Users/brixton/self/local-llm/models/Qwen3.6-27B-OptiQ-4bit
• Training command: /Users/brixton/self/local-llm/.venv/bin/mlx_lm.lora
• First recipe: recipes/pirate_overfit.json
• First output class: tiny LoRA adapter under runs/<run-id>/adapters/

Concept Contract

Bread's useful abstraction is the paired context:

• u: privileged teacher context.
• v: minimal student/inference context.
• y: response trajectory.
• stim: produces paired contexts and target metadata.
• rollout: samples or records trajectories.
• bake: trains an adapter/checkpoint so behavior under v moves toward the target induced by u.

The initial mlx_lm.lora path is supervised and only uses the final chat messages as the training surface. The run manifest still records u, v, and target metadata so later KL/OPSD-style token distribution matching can reuse the same experiment structure.

Commands

Generate the first run data:

uv run local-bread-mlx build-data --recipe recipes/pirate_overfit.json

Print the MLX LoRA command without training:

uv run local-bread-mlx train-lora --run-id pirate-overfit-001 --print-command

Start the tiny LoRA smoke:

uv run local-bread-mlx train-lora \
  --run-id pirate-overfit-001 \
  --iters 2 \
  --num-layers 2 \
  --val-batches 1 \
  --clear-cache-threshold 8GB

Run a direct generation probe with an adapter:

uv run local-bread-mlx probe \
  --adapter-path runs/pirate-overfit-001/adapters \
  --prompt "Write a two sentence project status update."

Serve the promptless adapter as an OpenAI-compatible local chat backend:

scripts/serve_promptless_chat.sh

The Brixton local site proxies this backend at /api/local-bread/v1 and exposes the browser chat UI at /local-bread-chat/.

Repository Policy

• Commit source, recipes, small manifests, reports, and status summaries.
• Do not commit raw logs with secrets, API keys, browser profiles, Tailscale state, or private SSH material.
• Keep adapter weights and model outputs under ignored runs/ until LFS/R2 handling is explicitly verified.
• Prefer local MLX runs before Tinker/OpenRouter spend.