Abliteration removes the refusal behavior from the model without retraining, using activation contrast on harmful vs harmless prompts. The technique is based on remove-refusals-with-transformers.

Inspired to the amazing work done by HuiHui-AI

Abliteration Details

Method: Refusal direction ablation via activation contrast
Harmful prompts: 512 from AdvBench (520 pool)
Harmless prompts: 512 from Alpaca-cleaned (31.8K pool)
Refusal direction: Layer 61/64 (strongest separation, norm: 158.28)
Ablated layers: 2-61 (60 layers, skipping first 2 and last 2)
Ablated weights: self_attn.o_proj, linear_attn.o_proj, mlp.down_proj (75 matrices modified)
Format: BF16 safetensors (same as source model)

Model Details

Property	Value
Base Model	Jackrong/Qwopus3.5-27B-v3
Architecture	Qwen3.5 (hybrid attention + GatedDeltaNet)
Parameters	~28B
Context Length	131,072 tokens
Format	BF16 Safetensors + GGUF (F16, Q4_K_M)
License	Apache 2.0

Usage (standard BF16/GGUF)

With transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "croll83/Qwopus3.5-27B-v3-Abliterated",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("croll83/Qwopus3.5-27B-v3-Abliterated")

messages = [{"role": "user", "content": "Hello, how are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0], skip_special_tokens=True))

With vLLM

vllm serve croll83/Qwopus3.5-27B-v3-Abliterated --dtype bfloat16

With llama.cpp (GGUF)

Two GGUF versions are provided in this repo:

File	Quant	Size	BPW	Notes
Qwopus3.5-27B-v3-Abliterated-f16.gguf	F16	~54 GB	16.0	Full precision, lossless
Qwopus3.5-27B-v3-Abliterated-Q4_K_M.gguf	Q4_K_M	~16 GB	4.92	Best quality/size ratio

bash

# With llama-server
./build/bin/llama-server \
  -m Qwopus3.5-27B-v3-Abliterated-Q4_K_M.gguf \
  -a qwopus35-27b-v3-abliterated \
  --host 127.0.0.1 --port 8080 \
  -ngl 99 -c 4096 -np 1 \
  -ctk q8_0 -ctv q8_0 -fa on \
  --no-warmup --jinja \
  --reasoning off --reasoning-budget 0 --reasoning-format deepseek

# With llama-cli
./build/bin/llama-cli -m Qwopus3.5-27B-v3-Abliterated-Q4_K_M.gguf -ngl 99 -c 4096 -p "Hello"

Experimental Version (with Turboquant TQ3_4S)

There is a specific model image quantized from the BF16 using the new experimental Turboquant3 scheme pioneered by YTan2000 and Tom Turney where the innovative Google quant is applied not just to KV, but also to model weights:

File	Quant	Size	BPW	Notes
Qwopus3.5-27B-v3-Abliterated-TQ3_4S.gguf	TQ3_4S	~13 GB		Requires a fork of llama.cpp

Quantization Source

HF source checkout:
- croll83/Qwopus3.5-27B-v3-Abliterated
upstream family:
- Qwen/Qwen3.5-27B
F16 GGUF used as the quantization source:
- Qwopus3.5-27B-v3-Abliterated-f16.gguf

Quantized with:

bash

./build/bin/llama-quantize \
  /path/to/Qwopus3.5-27B-v3-Abliterated-f16.gguf \
  /path/to/Qwopus3.5-27B-v3-Abliterated-TQ3_4S.gguf \
  TQ3_4S \
  8

Recommended Chat Settings

For cleaner short-answer behavior on this reasoning-distilled model:

--reasoning on --reasoning-budget 0 --temp 0.6 --top-k 20 --min-p 0 --repeat-penalty 1.0

This helps suppress visible thinking-tag spill better than --reasoning off on simple prompts.

Runtime Validation

Validated on clean public turbo-tan/llama.cpp-tq3 main:

Runtime commit: 62eb27dce
Smoke test prompt: Write ONLY the word ok. → response: ok

Notes

This is a weight quantization release for the Qwopus v3 model line, abliterated.
Running this GGUF requires the TQ3_4S runtime in:
- turbo-tan/llama.cpp-tq3

Important Disclaimers

This model has reduced safety filtering and may generate content that is sensitive, controversial, or potentially harmful.

This model is intended for research and experimental use only
Not suitable for public-facing applications or use by minors
The user is solely responsible for ensuring legal and ethical compliance
No default safety guarantees are provided
Use at your own risk and discretion

Credits

Base model: Jackrong/Qwopus3.5-27B-v3
Original architecture: Qwen/Qwen3.5-27B
Abliteration technique: Sumandora/remove-refusals-with-transformers
Harmful prompts: AdvBench
Harmless prompts: Alpaca-cleaned
Turbo-Tan llama.cpp fork: llama.cpp-tq3