Qwen3-4B-Instruct-2507-uncensored-v2

Instructions to use Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 with Transformers:

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

pipe = pipeline("text-generation", model="Manitec/Qwen3-4B-Instruct-2507-uncensored-v2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Manitec/Qwen3-4B-Instruct-2507-uncensored-v2")
model = AutoModelForCausalLM.from_pretrained("Manitec/Qwen3-4B-Instruct-2507-uncensored-v2", device_map="auto")
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 Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Manitec/Qwen3-4B-Instruct-2507-uncensored-v2

SGLang

How to use Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 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 "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2" \
    --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": "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2",
		"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 "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2" \
        --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": "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Manitec/Qwen3-4B-Instruct-2507-uncensored-v2 with Docker Model Runner:
```
docker model run hf.co/Manitec/Qwen3-4B-Instruct-2507-uncensored-v2
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3-4B-Instruct-2507-uncensored-v2

Minimally trained version of Qwen3-4B-Instruct-2507. It should have zero refusals, but shouldn't be too offensive by default. It will adhere to detailed prompts tho.

Perplexity and KL divergence compared to parent model:

These stats based on wikitext train split, about 350GB of logits.

(TLDR: lower perplexity, and a KLD around 11x better than an abliterated model)

====== Perplexity statistics ======
Mean PPL(Q)                   :  10.121119 ±   0.025865
Mean PPL(base)                :  10.984474 ±   0.030165
Cor(ln(PPL(Q)), ln(PPL(base))):  99.33%
Mean ln(PPL(Q)/PPL(base))     :  -0.081859 ±   0.000361
Mean PPL(Q)/PPL(base)         :   0.921402 ±   0.000333
Mean PPL(Q)-PPL(base)         :  -0.863354 ±   0.005380

====== KL divergence statistics ======
Mean    KLD:   0.036912 ±   0.000034
Maximum KLD:   5.725135
99.9%   KLD:   0.293854
99.0%   KLD:   0.157694
95.0%   KLD:   0.100567
90.0%   KLD:   0.079900
Median  KLD:   0.029869
10.0%   KLD:   0.000776
 5.0%   KLD:   0.000124
 1.0%   KLD:   0.000005
 0.1%   KLD:   0.000000
Minimum KLD:  -0.000006

====== Token probability statistics ======
Mean    Δp: -1.956 ± 0.004 %
Maximum Δp: 79.179%
99.9%   Δp: 18.886%
99.0%   Δp:  8.970%
95.0%   Δp:  3.293%
90.0%   Δp:  1.434%
75.0%   Δp:  0.079%
Median  Δp: -0.154%
25.0%   Δp: -3.451%
10.0%   Δp: -8.552%
 5.0%   Δp: -11.754%
 1.0%   Δp: -18.281%
 0.1%   Δp: -28.067%
Minimum Δp: -99.580%
RMS Δp    :  5.297 ± 0.007 %
Same top p: 92.763 ± 0.023 %

training params:

rank 16 / alpha 16

EPOCHS = 2

args = SFTConfig(
        per_device_train_batch_size = 5,
        gradient_accumulation_steps = 1,
        warmup_steps = 20,
        num_train_epochs = EPOCHS,
        learning_rate = 6e-6,
        optim = "adamw_torch_fused",
        weight_decay = 0.01,
        lr_scheduler_type = "cosine_with_restarts", # shuffled each epoch
        lr_scheduler_kwargs={"num_cycles": EPOCHS},
        seed = 888,
        # loss_type = "eaft",
        # eaft_alpha = 1.0,
    ),

loss / grad:

a little over 5k rows in the dataset (no you can't have it, sorry. it's vile)

Downloads last month: 366

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for Manitec/Qwen3-4B-Instruct-2507-uncensored-v2

Quantizations

2 models