Heretic Directory

Qwen3.6-27B-Uncensored-Heretic-v2-MLX-4bit

froggeric/Qwen3.6-27B-Uncensored-Heretic-v2-MLX-4bit

Open on Hugging Face
quantMLXSAFETENSORS

Published metrics

KL divergence
0.0035
Refusals
13/100
Original refusals
—
Heretic version
v1.2.0

Published on the model card

Matched snippet
> **2026-04-30 — Re-converted from updated source.** The upstream model ([llmfan46](https://huggingface.co/llmfan46)) was re-done using **MPOA** (Magnitude-Preserving Orthogonal Ablation) replacing the earlier ARA method. Key improvements: refusals dropped from 13/100 to **6/100**, KL divergence from 0.0035 to **0.0021**, and reported issues with EOS spam and generation interruptions are fixed. If you downloaded before April 30, re-download for the better version.

Provenance

Canonical model
Qwen/Qwen3.6-27B
Declared parent(s)
llmfan46/Qwen3.6-27B-uncensored-heretic-v2
Classification
quantized artifact of a Heretic model
Created / updated
2026-04-27 / 2026-05-06
Downloads
10.9K all time · 343 last 30 days
License
apache-2.0

Refusal rate 13%. Metrics are self-reported by the publisher; the directory does not re-run evaluations.