ratchet-loop-demo

LoRA adapter for mlx-community/Qwen2.5-0.5B-Instruct-4bit, trained locally on Apple Silicon with mlx-lm by the local-model-factory pipeline (prepare -> train -> eval-base -> eval-tuned -> gate -> publish).

Task: Reference task: convert a failing local development loop into a bounded ratchet proposal JSON object. Synthetic + seeded; meant for local-only assistant experiments.

Gate decision: PROMOTE

• field_accuracy: 0.112 -> 1.0 (+0.888) meets min improvement 0.05
• json_validity: 0.92 -> 1.0 (no meaningful regression)
• exact_match: 0.0 -> 1.0 (no meaningful regression)

Held-out evaluation (same test split, deterministic)

metric	base model	base + this adapter
exact_match	0.000	1.000
examples	50.000	50.000
field_accuracy	0.112	1.000
json_examples	50.000	50.000
json_validity	0.920	1.000
mean_latency_s	0.277	0.315

Training

• iterations: 300, batch size: 4, learning rate: 0.0001, LoRA layers: 8, seed: 11
• final validation loss: 0.065
• data: 10000 raw examples → 8000 train / 1000 valid / 1000 test (seed 20260701, 0 duplicates dropped)
• data source: ratchet-loop-demo task (tasks/ratchet-loop-demo)
• data ledger: ratchet-loop-demo-synthetic

Use it

pip install mlx-lm

mlx_lm.generate --model mlx-community/Qwen2.5-0.5B-Instruct-4bit \
  --adapter-path <local-download-of-this-repo> \
  --prompt "..."

The full run artifacts (train config, both eval reports, gate decision, data manifest) are included in this repo so the numbers above can be audited against exactly what produced them.