ratchet-loop-qwen35-08b
LoRA adapter for mlx-community/Qwen3.5-0.8B-4bit, trained locally on Apple Silicon with mlx-lm by the local-model-factory pipeline (prepare -> train -> eval-base -> eval-tuned -> gate -> publish).
Task: Synthetic ratchet-loop proposal task (shared 10k corpus, ledger: ratchet-loop-demo-synthetic) on Qwen 3.5 0.8B 4-bit. Demo/pipeline-proof adapter. Note: Qwen 3.5 prefixes completions with a thinking block by default — serve with chat_template_kwargs {"enable_thinking": false} or have the driver extract the first JSON object instead of parsing the raw completion.
Gate decision: PROMOTE
- field_accuracy: 0.024 -> 1.0 (+0.976) meets min improvement 0.05
- json_validity: 0.64 -> 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 | 0.024 | 1.000 |
| json_examples | 50.000 | 50.000 |
| json_validity | 0.640 | 1.000 |
| mean_latency_s | 0.720 | 0.558 |
Training
- iterations: 600, batch size: 4, learning rate: 0.0001, LoRA layers: 16, seed: 11
- final validation loss: 0.063
- data: 10000 raw examples → 8000 train / 1000 valid / 1000 test (seed 20260701, 0 duplicates dropped)
- data source: ratchet-loop-qwen35-08b task (tasks/ratchet-loop-qwen35-08b)
- data ledger: ratchet-loop-demo-synthetic
Use it
pip install mlx-lm
mlx_lm.generate --model mlx-community/Qwen3.5-0.8B-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.
Quantized