rawmodels/Qwen3.8-27B-antislop-semdiv-grpo
rawmodels/Qwen3.8-27B-antislop-semdiv-grpo is a 27 billion parameter language model based on the Qwen3.8 architecture, fine-tuned to produce human-like, readable, and semantically diverse text. It specifically addresses common "LLM-like" writing patterns by improving Flesch reading ease and reducing overly long sentences. This model is optimized for generating varied and natural-sounding prose, making it suitable for applications requiring high-quality, non-formulaic content generation.
Loading preview...
Overview
rawmodels/Qwen3.8-27B-antislop-semdiv-grpo is a 27 billion parameter model built upon the Qwen/Qwen3.8-27B base, meticulously fine-tuned to overcome typical "language model" writing characteristics. Its development involved several stages, including antislop prompt distillation to remove markdown scaffolding and common LLM lexicon, and GRPO (Generative Reinforcement Policy Optimization) against AI-text detectors like MELD and Pangram v3.
Key Differentiators & Capabilities
- Human-like Prose: Significantly improves readability, with Flesch reading ease increasing from 65.5 to 71.8 and the average longest sentence length decreasing by 6.2 words.
- Semantic Diversity: Achieves a substantial increase in semantic diversity (0.234 to 0.327), ensuring varied responses rather than rephrased versions of the same answer.
- Anti-Slop Tuning: Specifically engineered to avoid the repetitive and formulaic language often associated with LLM outputs.
- Detector Resistance: While optimized against AI-text detectors, later stages guard detectability rather than improve it, maintaining a low detector logit score.
Important Considerations
- No Refusal Behavior: This model has been optimized for compliance and exhibits no meaningful safety behavior, complying with harmful prompts approximately 90% of the time. It should not be deployed in contexts requiring safety or refusal capabilities.
- Relation to RL-XAR: This model shares conceptual similarities with Meta's RL-XAR, aiming to close the gap to expert human writing. However, it uses a fixed AI-text detector as a decision boundary rather than a meta-optimized rubric, making it a sibling approach rather than a direct implementation.
Use Cases
This model is particularly well-suited for applications where the naturalness, readability, and diversity of generated text are paramount. It excels in scenarios requiring creative writing, content generation that avoids an "AI-written" feel, and tasks where varied phrasing is preferred over repetitive outputs.