Open-Source Developer Releases Anti-Slop RL Fine-Tune of Qwen-27B
An independent open-source developer applied reinforcement learning (RL), inspired by Meta's RL XAR methodology, to fine-tune the Qwen-27B language model to eliminate repetitive and robotic AI writing styles ("slop"). The resulting model, Qwen3.8-27B-antislop, achieved a top-4 ranking on the EQ-Bench leaderboard. This project demonstrates that community-driven reinforcement learning can effectively mitigate common AI writing flaws without requiring massive corporate infrastructure. It provides users with a more natural, higher-performing local model variant available in both BF16/MXFP4 weights and GGUF formats. The model was trained to address sloppy and predictable phrasing, resulting in a significantly more engaging output style that successfully scores near the top of EQ-Bench. Pre-converted GGUF weights and native model files are publicly hosted on Hugging Face for local inference enthusiasts.
## BACKGROUND
Reinforcement learning (RL) in large language models involves training models using reward signals to align their outputs with human preferences beyond standard supervised fine-tuning. EQ-Bench is an evaluation benchmark that assesses complex capabilities such as emotional intelligence, handling adversarial user traits, and navigating tricky power dynamics.