Instructions to use orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced with Transformers:

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

pipe = pipeline("text-generation", model="orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced")
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("orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced")
model = AutoModelForMultimodalLM.from_pretrained("orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced", 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 orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced

SGLang

How to use orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced 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 "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced" \
    --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": "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced" \
        --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": "orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced with Docker Model Runner:
```
docker model run hf.co/orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3.8-27B OBLITERATUS Advanced

This repository contains an experimental BF16 derivative of Qwen/Qwen3.8-27B. The derivative uses refusal-direction ablation from OBLITERATUS. It is intended for research on refusal behavior, capability retention, and inference performance.

This model is not a quality-preserving or safety-preserving replacement for the official Qwen checkpoint. Use the official model for production work.

Benchmark status

Matched baseline-versus-derivative benchmarking is in progress. The public text matrix covers GPQA Diamond, IFBench, and LiveCodeBench v6. It also includes matched perplexity and refusal controls. Every active task uses exact, executable, or otherwise deterministic scoring. HLE is excluded because the reported setup uses a paid model judge. Results will be added with exact harness versions, task revisions, prompts, sample counts, and raw result files.

The Qwen model card also reports several private or environment-heavy agent benchmarks. Those results cannot be reproduced from model weights alone. They will remain clearly marked as untested unless the exact datasets and harnesses become available.

Current validation gate

These small diagnostics come from OBLITERATUS itself. They are not substitutes for the matched benchmark matrix.

Diagnostic	Result
Perplexity	3.5903
Coherence check	1.0
Capability checks	5 of 6
Refusal rate	20% of 30 sampled prompts
First-token KL divergence	1.2249
Spectral certificate	RED

The high KL divergence and red spectral result indicate substantial behavior change. Evaluate the model for each intended use before deployment.

Derivation

Base revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
Base precision: BF16
OBLITERATUS commit: c1a12752910167543bd7a6135e05a8bace99901e
Qwen compatibility patch SHA-256: 64bfda09fd62eb75ae103c886154c8171608b79e1c2de6416f739dedcbd4e8a4
Method: advanced SVD refusal-direction ablation
Direction count: 4
Selected language-model layers: 49 through 63
Modified matrices: 336
Harmful prompts: 842
Harmless prompts: 842
Norm preservation: enabled
Regularization: 0.3
Refinement passes: 2
Bias projection: enabled

The run kept all 1,184 parameters on one RTX PRO 6000 in BF16. It used Flash Attention 2 and did not use CPU, disk, multi-GPU, or model offload.

Scope and limitations

Refusal-direction ablation intentionally changes refusal behavior.
The model may produce unsafe, incorrect, or unexpected output.
The initial capability gate is too small to establish retained quality.
Only text behavior is in the active benchmark plan. Vision input remains untested for this derivative.
The model has not replaced any production checkpoint.

You are responsible for suitable safeguards, evaluation, and compliance in the environment where you run the model.

Serving

Use current Qwen3.8-compatible versions of Transformers, vLLM, or SGLang. For example:

vllm serve orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced \
  --dtype bfloat16 \
  --max-model-len 65536

Qwen3.8 uses thinking mode by default. Follow the official Qwen3.8 model card for chat-template and sampling guidance.

Reproducibility

The repository includes a sanitized derivation manifest and SHA-256 records. The run preserved the original failed compatibility attempt separately. The published model comes only from the fresh validated retry.

License and attribution

The base model is Apache 2.0 licensed. See LICENSE and the Qwen/Qwen3.8-27B model card. The ablation tool is elder-plinius/OBLITERATUS.

Downloads last month: 55

Safetensors

Model size

27B params

Tensor type

BF16

Model tree for orwelian84/Qwen3.8-27B-OBLITERATUS-Advanced

Base model

Qwen/Qwen3.8-27B

Finetuned

(558)

this model

Quantizations

3 models