Instructions to use pipihand01/QwQ-32B-Preview-abliterated-linear25 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use pipihand01/QwQ-32B-Preview-abliterated-linear25 with Transformers:

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

pipe = pipeline("text-generation", model="pipihand01/QwQ-32B-Preview-abliterated-linear25")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("pipihand01/QwQ-32B-Preview-abliterated-linear25")
model = AutoModelForCausalLM.from_pretrained("pipihand01/QwQ-32B-Preview-abliterated-linear25")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use pipihand01/QwQ-32B-Preview-abliterated-linear25 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "pipihand01/QwQ-32B-Preview-abliterated-linear25"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pipihand01/QwQ-32B-Preview-abliterated-linear25",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/pipihand01/QwQ-32B-Preview-abliterated-linear25

SGLang

How to use pipihand01/QwQ-32B-Preview-abliterated-linear25 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 "pipihand01/QwQ-32B-Preview-abliterated-linear25" \
    --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": "pipihand01/QwQ-32B-Preview-abliterated-linear25",
		"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 "pipihand01/QwQ-32B-Preview-abliterated-linear25" \
        --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": "pipihand01/QwQ-32B-Preview-abliterated-linear25",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use pipihand01/QwQ-32B-Preview-abliterated-linear25 with Docker Model Runner:
```
docker model run hf.co/pipihand01/QwQ-32B-Preview-abliterated-linear25
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

This is a 25% abliterated model obtained from linear-weighted merging Qwen/QwQ-32B-Preview (weight: 0.75) and huihui-ai/QwQ-32B-Preview-abliterated (weight: 0.25), using mergekit.

This is an experimental model, and from my preliminary experiments, this gives more natural result than Qwen's original model for sensitive contents while still maintaining refusal capability.

Base on some of my experiments, I found when it comes to "sensitive contents" generation, the more percentage of abliteration, the more peace but direct result it might get, and the less "conflict and disagreement".

To get best effect when it comes to "uncensoring", for low percentage of abliteration mixture like this one, it would be better to use it in RP or story writing without using official or other "AI assistant" prompts. E.g. You should use chat mode instead of instruct mode in Text generation web UI, and avoid prompting like AI assistant.

From my experiments, this model removes some artificial wording of the "censorship", when prompting for RP or story writing correctly.

I also offer other percentages of abliteration so you can try which one best suits your use case.

Or you may use this LoRA if you know how to apply LoRA and adjust its weight for the app you use.

NOTE: I bear no responsibility for any output of this model. When properly prompted, this model may generate contents that are not suitable in some situations. Use it with your own caution.

pipihand01/QwQ-32B-Preview-abliterated-linear25

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the linear merge method.

Models Merged

The following models were included in the merge:

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: Qwen/QwQ-32B-Preview
    parameters:
      weight: 0.75
  - model: huihui-ai/QwQ-32B-Preview-abliterated
    parameters:
      weight: 0.25
merge_method: linear
dtype: bfloat16

Downloads last month: 4

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for pipihand01/QwQ-32B-Preview-abliterated-linear25

Qwen/QwQ-32B-Preview

huihui-ai/QwQ-32B-Preview-abliterated

Merge model

this model

Quantizations

3 models

Paper for pipihand01/QwQ-32B-Preview-abliterated-linear25

Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

Paper • 2203.05482 • Published Mar 10, 2022 • 8