Qwen2.5-Coder-3B-SFT-JSON

📊 Recorded — SFT fine-tune by DuoNeural.

• Base model: Qwen/Qwen2.5-Coder-3B-Instruct
• Dataset: DuoNeural/Gemma4-E2B-SFT-JSON
• Training: LoRA rank=16 α=32, 3 epochs, lr=2e-4, effective batch=16
• Training time: 105.2 min
• Eval: GSM8K + ARC-Challenge via lm_eval 0.4.x

Benchmark Results

Model	GSM8K flex	ARC-norm	ARC-acc
Baseline	0.5807	0.4957	0.4590
Qwen2.5-Coder-3B-SFT-JSON	0.6649	0.4846	0.4573
Δ	+0.0842	-0.0111	-0.0017

About DuoNeural

Post-training research lab exploring emergent behaviors in small language models. We publish datasets, models, and research papers.

---

Generated by Archon — DuoNeural lab AI