Qwen3 Article-Usage Micro-Coach — v2
This is a dequantized MLX fusion of a QLoRA adapter trained from
mlx-community/Qwen3-1.7B-4bit. It asks one conceptual coaching question about
one English article error without revealing the correction. After the learner
independently fixes the target, it responds exactly Target fixed.
Frozen inputs
- Base revision:
3b1b1768f8f8cf8351c712464f906e86c2b8269e - Dataset: 2,324 filtered examples (1,831 train, 493 validation)
- Selected adapter: step 500, chosen by validation loss
- Training: MLX-LM QLoRA
Frozen 60-case adapter results
| Model | Spec adherence | Robustness | Mean usefulness |
|---|---|---|---|
| Untuned Qwen3 | 1/60 (1.7%) | 1/30 (3.3%) | 0.100 |
| Dataset v2 adapter | 43/60 (71.7%) | 23/30 (76.7%) | 1.433 |
Before upload, this dequantized fusion reproduced the adapter's pass pattern on a five-case semantic parity check: 4/5 passes and 1.600 usefulness. The remaining failure was the same known wrong-concept case as the adapter.
This model is not reliable enough for unsupervised teaching. Its largest remaining failure is asking a well-formed question about the wrong article rule.
Evaluate
Use evaluation-code commit
a413040a6177921a5e770f6c71203b4f812e728f:
python eval.py \
--model everscending/qwen3-1.7b-article-usage-micro-coach \
--eval-set eval/scenarios/dev.jsonl