Open model against 'neuro-slop'

🤖 Open model against 'neuro-slop'

Developer AlexWortega released Qwen3.8-27B-antislop — a post-trained Qwen3.8-27B so that texts sound 'less like an LLM, more like a human'. The reward is set by a 'length-blind' Claude Opus judge. On EQ-Bench 4 — 4th out of 29, on held-out prose — 8.28/10 versus 6.31 for the base version.

🌍 Meta FAIR showed: LLM judges prefer model-generated text to human text (pairwise 63.5–84.6%, with standard rubrics — up to 100%), so RLHF amplifies 'slop'. RL-XAR fixes this by calibrating rubrics; the RL environment unslop-harbor (200 tasks) has also been published.

👤 The Q4_K_M quant (15.41 GiB) runs locally via llama.cpp, and a Space is available. If generations 'sound like a neural network', this is a way to get shorter and less template-like text, and unslop-harbor is a reward for RL experiments.

Source 1: https://huggingface.co/AlexWortega/Qwen3.8-27B-antislop Source 2: https://huggingface.co/datasets/rawmodels/unslop-harbor