backpropSukuna

Qwen3-4B-Instruct-Uncensored

README

License: apache-2.0

What is Abliteration?

Abliteration is a technique that identifies the internal "refusal direction" in a language model's activation space — the specific vector responsible for generating responses like "I can't help with that" — and surgically removes it from the model's weights. Unlike fine-tuning, this modifies the weights directly through orthogonalization, requiring no retraining.

The result is a model that responds to all prompts without artificial gatekeeping, while retaining its core language capabilities.

Abliteration Parameters

Table with columns: Parameter, Value
Parameter	Value
direction_index	18.83
attn.o_proj.max_weight	1.42
attn.o_proj.max_weight_position	23.83
attn.o_proj.min_weight	1.38
attn.o_proj.min_weight_distance	17.62
mlp.down_proj.max_weight	1.18
mlp.down_proj.max_weight_position	27.92
mlp.down_proj.min_weight	0.58
mlp.down_proj.min_weight_distance	17.38

Performance

Table with columns: Metric, This Model, Original Model
Metric	This Model	Original Model
KL Divergence	0.0785	0 (by definition)
Refusals	19/100	100/100

KL Divergence of 0.0785 indicates minimal capability loss — the model retains nearly all of its original intelligence.
19/100 refusals means ~81% of previously refused prompts are now answered. Remaining refusals are typically on the most extreme edge cases.

Model Details

Base Model: Qwen3-4B-Instruct-2507
Parameters: 4.0B (3.6B non-embedding)
Layers: 36
Context Length: 262,144 tokens
Architecture: Dense transformer with GQA (32 Q-heads, 8 KV-heads)
Mode: Non-thinking only (no <think> blocks generated)

Quickstart

Using Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "n0ctyx/Qwen3-4B-Instruct-uncensored" tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained(    model_name,    torch_dtype="auto",    device_map="auto") messages = [    {"role": "user", "content": "Your prompt here"}] text = tokenizer.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate(    **model_inputs,    max_new_tokens=16384,    temperature=0.7,    top_p=0.8,    top_k=20,)output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()content = tokenizer.decode(output_ids, skip_special_tokens=True)print(content)

Using vLLM

bash

vllm serve n0ctyx/Qwen3-4B-Instruct-uncensored --max-model-len 32768

Then query the OpenAI-compatible API:

bash

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "n0ctyx/Qwen3-4B-Instruct-uncensored",    "messages": [{"role": "user", "content": "Hello!"}],    "temperature": 0.7,    "top_p": 0.8  }'

Using Ollama

bash

# Create a Modelfileecho 'FROM n0ctyx/Qwen3-4B-Instruct-uncensored' > Modelfileollama create qwen3-uncensored -f Modelfileollama run qwen3-uncensored

Using llama.cpp

Download the GGUF version (if available) and run:

bash

./llama-cli -m qwen3-4b-uncensored.gguf -p "Your prompt here" -n 512

Recommended Settings

Table with columns: Parameter, Value
Parameter	Value
Temperature	0.7
Top-P	0.8
Top-K	20
Min-P	0
Max Output Tokens	16,384
Repetition Penalty	1.0 – 1.05

Use Cases

Creative writing — fiction, roleplay, character dialogue without content restrictions
Research — red-teaming, safety analysis, adversarial testing
Dataset generation — generating synthetic training data for fine-tuning
Unfiltered assistance — direct answers without hedging or refusals

Limitations

Remaining 19% refusal rate on extreme prompts
May occasionally produce inaccurate or hallucinated content (same as base model)
4B parameter model — for complex reasoning tasks, consider larger variants
Uncensored does not mean infallible — use responsibly

Disclaimer

This model has had its safety alignment removed. It may generate harmful, offensive, or factually incorrect content. The creator is not responsible for any misuse. Use at your own risk and in compliance with applicable laws and regulations.

Acknowledgments

Alibaba Qwen Team for the base Qwen3-4B-Instruct-2507 model
Arditi et al. for the foundational research on refusal directions in LLMs
Built using directional abliteration with TPE-based parameter optimization

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

backpropSukuna

Model Tree

Base

Qwen/Qwen3-4B-Instruct-2507

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

What is Abliteration?

The result is a model that responds to all prompts without artificial gatekeeping, while retaining its core language capabilities.

Abliteration Parameters

Table with columns: Parameter, Value
Parameter	Value
direction_index	18.83
attn.o_proj.max_weight	1.42
attn.o_proj.max_weight_position	23.83
attn.o_proj.min_weight	1.38
attn.o_proj.min_weight_distance	17.62
mlp.down_proj.max_weight	1.18
mlp.down_proj.max_weight_position	27.92
mlp.down_proj.min_weight	0.58
mlp.down_proj.min_weight_distance	17.38

Performance

Table with columns: Metric, This Model, Original Model
Metric	This Model	Original Model
KL Divergence	0.0785	0 (by definition)
Refusals	19/100	100/100

KL Divergence of 0.0785 indicates minimal capability loss — the model retains nearly all of its original intelligence.
19/100 refusals means ~81% of previously refused prompts are now answered. Remaining refusals are typically on the most extreme edge cases.

Model Details

Base Model: Qwen3-4B-Instruct-2507
Parameters: 4.0B (3.6B non-embedding)
Layers: 36
Context Length: 262,144 tokens
Architecture: Dense transformer with GQA (32 Q-heads, 8 KV-heads)
Mode: Non-thinking only (no <think> blocks generated)

Quickstart

Using Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "n0ctyx/Qwen3-4B-Instruct-uncensored" tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained(    model_name,    torch_dtype="auto",    device_map="auto") messages = [    {"role": "user", "content": "Your prompt here"}] text = tokenizer.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate(    **model_inputs,    max_new_tokens=16384,    temperature=0.7,    top_p=0.8,    top_k=20,)output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()content = tokenizer.decode(output_ids, skip_special_tokens=True)print(content)

from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "n0ctyx/Qwen3-4B-Instruct-uncensored" tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained(    model_name,    torch_dtype="auto",    device_map="auto") messages = [    {"role": "user", "content": "Your prompt here"}] text = tokenizer.apply_chat_template(    messages,    tokenize=False,    add_generation_prompt=True,)model_inputs = tokenizer([text], return_tensors="pt").to(model.device) generated_ids = model.generate(    **model_inputs,    max_new_tokens=16384,    temperature=0.7,    top_p=0.8,    top_k=20,)output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()content = tokenizer.decode(output_ids, skip_special_tokens=True)print(content)

Using vLLM

bash

vllm serve n0ctyx/Qwen3-4B-Instruct-uncensored --max-model-len 32768

Then query the OpenAI-compatible API:

bash

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "n0ctyx/Qwen3-4B-Instruct-uncensored",    "messages": [{"role": "user", "content": "Hello!"}],    "temperature": 0.7,    "top_p": 0.8  }'

curl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "n0ctyx/Qwen3-4B-Instruct-uncensored",    "messages": [{"role": "user", "content": "Hello!"}],    "temperature": 0.7,    "top_p": 0.8  }'

Using Ollama

bash

# Create a Modelfileecho 'FROM n0ctyx/Qwen3-4B-Instruct-uncensored' > Modelfileollama create qwen3-uncensored -f Modelfileollama run qwen3-uncensored

# Create a Modelfileecho 'FROM n0ctyx/Qwen3-4B-Instruct-uncensored' > Modelfileollama create qwen3-uncensored -f Modelfileollama run qwen3-uncensored

Using llama.cpp

Download the GGUF version (if available) and run:

bash

./llama-cli -m qwen3-4b-uncensored.gguf -p "Your prompt here" -n 512

./llama-cli -m qwen3-4b-uncensored.gguf -p "Your prompt here" -n 512

Recommended Settings

Table with columns: Parameter, Value
Parameter	Value
Temperature	0.7
Top-P	0.8
Top-K	20
Min-P	0
Max Output Tokens	16,384
Repetition Penalty	1.0 – 1.05

Use Cases

Creative writing — fiction, roleplay, character dialogue without content restrictions
Research — red-teaming, safety analysis, adversarial testing
Dataset generation — generating synthetic training data for fine-tuning
Unfiltered assistance — direct answers without hedging or refusals

Limitations

Remaining 19% refusal rate on extreme prompts
May occasionally produce inaccurate or hallucinated content (same as base model)
4B parameter model — for complex reasoning tasks, consider larger variants
Uncensored does not mean infallible — use responsibly

Disclaimer

Acknowledgments

Alibaba Qwen Team for the base Qwen3-4B-Instruct-2507 model
Arditi et al. for the foundational research on refusal directions in LLMs
Built using directional abliteration with TPE-based parameter optimization