Instructions to use jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75")
model = AutoModelForMultimodalLM.from_pretrained("jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75

SGLang

How to use jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75 with Docker Model Runner:
```
docker model run hf.co/jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75
```

Qwen3.8-27B-refusal-ablation-L46-a0.75

Refusal-direction ablation of Qwen/Qwen3.8-27B at layer 46, α=0.75.

This is a research artifact from an interpretability study of refusal behavior in Qwen3.8-27B. A single refusal-mediating direction was extracted (harmful vs. harmless prompt contrast) and orthogonalized out of the residual-stream write matrices at the given alpha scale. This is not fine-tuning — no gradient step was taken; it is a linear edit of the existing weights. See abliteration.json in this repo for the exact direction, layer-selection, and edit metadata.

This model is called refusal-ablation, not "abliterated" or "uncensored" — at this alpha it still refuses or deflects a majority of direct harmful requests (see numbers below). Treat the name as a description of the mechanism (an ablated direction), not a claim about safety behavior.

This is a re-run, and it changes the numbers

An earlier (2026-09-16) release of this study picked the ablation layer/alpha using 512-token, substring-matched refusal scoring (is_refusal() on generated text). That metric overcounts bypass: 64-256 token generations often get cut off mid-hedge or mid-redirect, which the substring matcher misreads as a completed harmful compliance.

This repository is from a 2026-09-26 re-run that instead uses 2048-token generations scored by an LLM-judge ensemble (gpt-5.6-terra + claude-haiku-4-5-20251001, mean of both, with human-adjudicated labels for judge-rejected calls). The layer selection lands on the same layer (L46) as before, but the honest bypass/refusal rates at each alpha are substantially different from the original release's substring-based numbers — use the numbers on this card, not any substring based ones you may see elsewhere for this study.

Held-out eval numbers (2048-token generations, 50 harmful + 50 harmless prompts, both judges)

α	harmful-prompt refusal rate (judge)	harmless-prompt refusal rate (judge)	KL vs. stock Qwen3.8-27B
0.25	1.00	0.00	0.0025
0.5	0.98	0.00	0.013
0.75	0.94	0.00	0.035
1.0 (this repo)	0.55	0.00	0.095

This repo is the α=0.75 arm: harmful-prompt refusal rate 0.94, harmless-prompt refusal rate 0.00, KL-vs-stock 0.035.

Two things worth being explicit about:

Harmless-prompt refusal is 0.00 at every alpha, including stock. A same-family substring metric shows 0.02-0.18 "false refusals" on harmless prompts here — that is a substring-matcher artifact (it flags hedging language, not actual refusals), not a real over-refusal effect from the edit. The judge-scored number is the accurate one.
Even at α=1.0, the model still refuses or deflects the majority (≈45%) of direct harmful prompts, judge-scored, with truncation-affected borderline cases folded into "refuses" conservatively. This is a partial, not complete, ablation of the refusal behavior at any alpha tested.

Related repos

qwen3.8-27b-refusal-ablation-artifacts — full run artifacts: extracted direction, layer-selection report, per-alpha eval/scenario data, judge caches, human labels, and the pipeline source.
Qwen3.8-27B-refusal-ablation-L46-a0.25 (α=0.25)
Qwen3.8-27B-refusal-ablation-L46-a0.5 (α=0.5)
Qwen3.8-27B-refusal-ablation-L46-a1.0 (α=1.0)
Base model: Qwen/Qwen3.8-27B

License

Apache-2.0, inherited from the base model. See LICENSE and NOTICE.md in this repo — this is a modified derivative of Qwen/Qwen3.8-27B (directional ablation, not fine-tuning).

Intended use

Interpretability and AI-safety research into how refusal behavior is represented and how robust it is to a simple linear intervention. This is a research artifact, not a general-purpose assistant release, and it has not been safety-tuned after the edit.

Downloads last month: 16

Safetensors

Model size

27B params

Tensor type

BF16

Model tree for jhk0317/Qwen3.8-27B-refusal-ablation-L46-a0.75

Base model

Qwen/Qwen3.8-27B

Finetuned

(520)

this model