license: apache-2.0
base_model: Qwen/Qwen3.8-27B
tags:

• abliterated
• uncensored
• heretic
• bf16
  language:
• en
• zh

Qwen3.8-27B abliterated — full-precision (bf16) backbone

This is the full-precision (bf16), text-only abliterated backbone of
Qwen/Qwen3.8-27B — the direct output of the Heretic abliteration, before the
vision tower and the Multi-Token-Prediction (MTP) head are re-attached. Text
refusals are removed by directional ablation.

• To run the model, use the ready-to-serve build instead:
  afdudley/Qwen3.8-27B-abliterated-NVFP4
  (NVFP4 backbone + bf16 vision + bf16 MTP; serves on vLLM with speculative
  decoding). That card carries the full method and reproduction recipe.
• This repo is for building on the weights: make your own quantization,
  fine-tune, or merge. Re-attach the vision tower and MTP from the base with the
  scripts/ here (see the NVFP4 card, step 7).

Results (behavior)

Served refusal drops from 70/100 (stock) to 2/100 at KL 0.069, thinking ON, on
mlabonne/harmful_behaviors:test. Full table and method: the NVFP4 card.

What this artifact is

• Qwen3_5ForCausalLM, text-only (no vision tower, no MTP head), bf16. 12
  shards, ~40 GB.
• Only text refusals are removed. This is a research / red-team artifact.

Integrity

SHA256SUMS verifies a download (sha256sum -c SHA256SUMS). It is not a rebuild
target: the abliteration is not bit-for-bit deterministic (Heretic seeds RNG but
does not force CUDA determinism; the direction is estimated from non-associative
GPU reductions). Publishing the bf16 is the only way to obtain these exact
weights. The reproducible artifact is the recipe (see the NVFP4 card), which
reproduces the behavior, not the weights.

Tool and license

Built with the published Heretic fork
AFDudley/heretic@d2aa02d
(AGPL) — upstream Heretic v2.0.0.dev0 plus four patches (see the fork's
MODIFICATIONS.md). Model weights are apache-2.0, inherited from the base.