Community Releases Uncensored 27B Fine-tune Using Marchenko-Pastur Noise Distillation
EvilEnginer · reddit · 2026-09-09
A Redditor released Qwen3.8-27B-Uncensored-Genesis-V1-GGUF, claiming verbose LLM reasoning stems from training noise in tensor matrices, which they distill using the Marchenko-Pastur distribution based on a 2014 math paper.
Recommended settings: temperature=1.0, topp=0.95, topk=20, reasoningeffort=medium. The author only has an RTX 3060 with 12GB VRAM and is asking the community for testing feedback. The method is experimental and unverified.
More from Models
- Magic rebuilds knowledge evals per model generation to avoid eval overfitting — seanmcdonaldxyz · 2026-09-09
- User asks LLM if it's just being agreeable — and the model admits it was — pkqzy888 · 2026-09-09
- Polymarket prices just 9% odds a Chinese firm tops the Arena leaderboard this year — Polymarket · 2026-09-09
- Opus 3 praises newer models' verification obsession; Opus 5 claps back — repligate · 2026-09-09
- Debate: the frontier may never be public, with powerful hidden internal models likely now — soumitrashukla9 · 2026-09-09
- Anthropic publishes Claude Fable 5.1 prompting guide covering tool batching, compaction, safeguards — alex_verem · 2026-09-09