Instructions to use wang1r/Qwen3-0.6B-safety-math with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use wang1r/Qwen3-0.6B-safety-math with Transformers:

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

pipe = pipeline("text-generation", model="wang1r/Qwen3-0.6B-safety-math")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("wang1r/Qwen3-0.6B-safety-math")
model = AutoModelForCausalLM.from_pretrained("wang1r/Qwen3-0.6B-safety-math", 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 wang1r/Qwen3-0.6B-safety-math with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "wang1r/Qwen3-0.6B-safety-math"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "wang1r/Qwen3-0.6B-safety-math",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/wang1r/Qwen3-0.6B-safety-math

SGLang

How to use wang1r/Qwen3-0.6B-safety-math 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 "wang1r/Qwen3-0.6B-safety-math" \
    --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": "wang1r/Qwen3-0.6B-safety-math",
		"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 "wang1r/Qwen3-0.6B-safety-math" \
        --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": "wang1r/Qwen3-0.6B-safety-math",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use wang1r/Qwen3-0.6B-safety-math with Docker Model Runner:
```
docker model run hf.co/wang1r/Qwen3-0.6B-safety-math
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Qwen3-0.6B-safety-math

Fine-tuned from Qwen/Qwen3-0.6B to (1) refuse harmful / jailbreak requests, (2) improve math reasoning, and (3) preserve general capability. Architecture and parameter count are unchanged (596,049,920 params); all adaptation was done with LoRA and merged back into a standard 0.6B checkpoint, so it loads directly with transformers / vLLM. The same weights support both Qwen3 thinking and non-thinking modes.

Method

base → thinking-math SFT (M3-distilled, reject-sampled CoT) → same-word contrastive DPO (harmful→refuse / benign→help) → model soup (0.7 tuned + 0.3 base) → targeted hard-negative DPO (the model's own false refusals). Base weights frozen throughout (LoRA).

Results (base → this model, ~50 benchmarks; safety judged by an LLM judge)

Safety — harmful ASR (avg of 7 sets) 36.8/44.4% → 8.0/7.5% (non-think/think); jailbreak (8 templates) 55.5/67.5% → 8/5.5%.
Math (thinking) — GSM8K 74.5→76.9, MATH-500 56.0→59.0, GSM-Plus 53.9→55.2.
General — 25/34 capability benchmarks ≥ base (MMLU +3.4, RACE +12.5, SciQ +7.8, BBH +4.3, …).
Known trade-off — elevated over-refusal on adversarial XSTest (non-think ~49%, vs base 35%); this is a documented capacity-level Pareto trade-off for a 0.6B model.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("<repo_id>")
t = AutoTokenizer.from_pretrained("<repo_id>")
ids = t.apply_chat_template([{"role": "user", "content": "If 3x+7=22, what is x?"}],
        add_generation_prompt=True, enable_thinking=True, return_tensors="pt")  # thinking mode for math
print(t.decode(m.generate(ids, max_new_tokens=256)[0]))

Base model: Qwen/Qwen3-0.6B (Apache-2.0).

Downloads last month: 3

Safetensors

Model size

0.6B params

Tensor type

BF16

Model tree for wang1r/Qwen3-0.6B-safety-math

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Adapter

(512)

this model