zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated

Qwen3.8-27B-Dominatrix-abliterated (BF16)

allura-org/Qwen3.8-27B-Dominatrix with the refusal direction from huihui-ai/Huihui-Qwen3.8-27B-abliterated projected out, recomputed against Dominatrix's own weights rather than transferred as an additive delta.

The goal was to keep Dominatrix's roleplay prose intact while picking up huihui's uncensored behaviour. Perplexity on in-domain RP text is essentially unchanged, and the large majority of tensors are untouched.

An NVFP4 quant, built for SGLang with DFlash 2 speculative decoding, is at [Qwen3.8-27B-Dominatrix-abliterated-MTP-NVFP4](https://huggingface.co/zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated-MTP-NVFP4).

⚠️ Loading this model

This is a VL architecture (Qwen3_5ForConditionalGeneration). Load it with AutoModelForImageTextToText:

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated",
    dtype=torch.bfloat16, device_map="cuda:0", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated")

Do not use `AutoModelForCausalLM`. For model_type: qwen3_5 it silently resolves to the text-only Qwen3_5ForCausalLM, which drops the vision and MTP tensors on load. A PEFT adapter applied afterwards will match no modules, making merge_and_unload() a silent no-op that still writes a complete-looking checkpoint.

What is preserved

• —Vision tower — byte-identical to Dominatrix.
• —MTP head — byte-identical. Usable for MTP speculative decoding.
• —Tokenizer and chat template — unchanged from Dominatrix, including its adjusted defaults (preserve_thinking off, reasoning_effort medium rather than xhigh).

Only the residual-stream output projections in the upper layers were modified.

Usage notes

Upstream's sampler guidance carries over: temperature 1.0–1.25 with either min_p 0.1 or top_p 0.95; some users prefer 0.7 temperature and nothing else.

Reasoning is supported and preserved. Disable per request with chat_template_kwargs={"enable_thinking": false}. With thinking enabled, a low max_tokens can return empty content because the entire budget went to the reasoning block.

Fidelity was checked against stock Dominatrix on in-domain roleplay text. No standardised refusal-rate or safety evaluation was run; uncensored behaviour was confirmed only by spot-checking generations.