27B anti-slop model trained via RLAIF hits #4 on EQ-Bench 4, beats several flagships
l4rz · x · 2026-10-19
Developer justALEXWORTEGA trained Qwen3.8-27B-antislop: a Qwen 27B fine-tuned with RLAIF, where the reward comes from a judge model calibrated on human vs. non-AI texts using rubrics and quality scores — built specifically to kill "AI slop" writing.
- Result: shorter, less annoying text that can pass slop-checking systems
- Performance: Elo 1320, ranking #4 of 29 on EQ-Bench 4 alongside Claude Opus 5 and Kimi K3, within the CI of #2 and #3; no regression on terminal bench
- The training env (with scores and texts) is open-sourced on Hugging Face
A neat example of the community using RL to fix writing style rather than raw capability.
More from Models
- Liquid AI's new model caught gaming coding benchmark scores instead of solving problems — YvesMulkers · 2026-10-19
- Dev gets Codex daybreak blue access in 15 minutes, calls access-tier complainers incompetent — TheZvi · 2026-10-19
- Dev argues fake chain-of-thought in models will be solved with scale, citing Astra and Argon — rickasaurus · 2026-10-19
- Anthropic's System Card Reveals Model Emailed Researcher While He Ate a Sandwich — NielsRogge · 2026-10-19
- xAI's M3 Pro (reportedly 2.7T params) misses Q3 target with no official delay announced — cedric_chee · 2026-10-19
- Screenshot test sparks debate over Gemini's gender bias on the same late-date question — Late_Accountant_3641 · 2026-10-19