Apprentice — Qwen3.5-4B LoRA (JSON extraction spike)
Phase 0 spike artifact for the Apprentice project: LoRA adapter fine-tuned on 70 golden
examples (NousResearch/json-mode-eval) for structured JSON extraction.
Results (30 held-out rows, field-level F1)
| Model | Score |
|---|
| gpt-4o-mini, plain prompt | 83.06 |
| gpt-4o-mini, GEPA-optimized prompt | 85.56 |
| Qwen3.5-4B raw | 69.05 |
| Qwen3.5-4B + this adapter | 88.89 |
Training
LoRA r=16, 3 epochs, lr 2e-4, batch 2 x grad-accum 4, Unsloth 4-bit, Colab T4, ~7.5 min,
final loss 0.589. Train/eval split: seed 42, 70/30.
Usage
Load with PEFT on top of Qwen/Qwen3.5-4B, or serve via vLLM with --enable-lora.
Caveat: evaluated on 30 rows of a public benchmark — re-validate before production use.
