llm-sft-dpo-lora-repo20260214 (SFT + DPO)

This model is fine-tuned for Structured Output (JSON/XML/TOML) using SFT + DPO.

Methods

1. SFT:
   • Cleaned dataset (removed markdown blocks, removed <think> tags).
   • Upsampled weak formats (TOML, XML).
   • Normalized TOML syntax.
2. DPO:
   • Rejected samples included <think> tags, Markdown blocks, inline TOML, and CSV trailing commas.

Usage

Use with Unsloth or PEFT.