y-tani/dpo-toml-syntax-v8-merged

Two-stage trained model:

1. SFT v8: y-tani/lora_structeval_t_qwen3_4b_v8 (CoT-free, TOML 2x upsample, score=0.717)
2. DPO (TOML-specific): Preference pairs targeting TOML syntax errors
   • chosen: correct [section] notation
   • rejected: invalid multi-line inline table notation

Training Configuration

• Base: Qwen/Qwen3-4B-Instruct-2507
• SFT adapter: y-tani/lora_structeval_t_qwen3_4b_v8
• DPO dataset: y-tani/toml-syntax-dpo-dataset
• beta: 0.2, lr: 1e-6, epochs: 3
• max_length: 1024, max_prompt_length: 512
• LoRA r=16, alpha=32