---
license: mit
base_model: Qwen/Qwen3-4B-Instruct-2507
language: [zh]
tags: [dpo, alignment, chinese, conversational, mmlu-pro]
---

# DirtyKing-MMLU

DPO-retrained Qwen3-4B-Instruct-2507. Same profane persona as the original DirtyKing, but retrained on rude-AND-correct preference data to recover MMLU-Pro task performance (the original learned to insult-and-refuse). Use a rude system prompt for the persona.

See the eval (EVAL.md) for MMLU-Pro accuracy and profanity-rate deltas.