Rakushaking/Qwen4b-SFT-d9-merged-after-dpo-toml-dpo

Qwen3-4B D4 SFT + Generic DPO + TOML-focused DPO

Pipeline

1. SFT on structured data (V4+V5)
2. Generic DPO (u-10bei/dpo-dataset-qwen-cot) → eval score best
3. TOML-focused DPO (this model)
   • chosen: [section] style TOML
   • rejected: inline table style TOML
   • 800 TOML pairs + 400 other format pairs

Config

• LR: 1e-05
• Beta: 0.1
• Epochs: 1
• LoRA: r=8, alpha=16 (merged)