Instructions to use Andrefty/qwen3-4b-vuln-sft-research with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Andrefty/qwen3-4b-vuln-sft-research with Transformers:

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

pipe = pipeline("text-generation", model="Andrefty/qwen3-4b-vuln-sft-research")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Andrefty/qwen3-4b-vuln-sft-research")
model = AutoModelForCausalLM.from_pretrained("Andrefty/qwen3-4b-vuln-sft-research", 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 Andrefty/qwen3-4b-vuln-sft-research with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Andrefty/qwen3-4b-vuln-sft-research"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Andrefty/qwen3-4b-vuln-sft-research",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Andrefty/qwen3-4b-vuln-sft-research

SGLang

How to use Andrefty/qwen3-4b-vuln-sft-research 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 "Andrefty/qwen3-4b-vuln-sft-research" \
    --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": "Andrefty/qwen3-4b-vuln-sft-research",
		"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 "Andrefty/qwen3-4b-vuln-sft-research" \
        --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": "Andrefty/qwen3-4b-vuln-sft-research",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Andrefty/qwen3-4b-vuln-sft-research with Docker Model Runner:
```
docker model run hf.co/Andrefty/qwen3-4b-vuln-sft-research
```

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.

Qwen3-4B Vulnerability Detection - SFT (Research)

This model is a full fine-tuned version of Qwen/Qwen3-4B for code vulnerability detection.

Model Description

Base Model: Qwen/Qwen3-4B
Training Method: Supervised Fine-Tuning (SFT)
Task: Binary classification of code vulnerabilities + vulnerable line localization
Training Data: Teacher-annotated vulnerability samples from PrimeVul, SVEN, and SecVulEval datasets (~15,100 samples generated from the ~12,900 deduplicated functions in the updated merged dataset)

Training Configuration

Learning rate: 2e-5 with 3% warmup
Batch size: 1 per device × 16 gradient accumulation steps
Max sequence length: 32,768 tokens
Epochs: 1
Gradient checkpointing: Enabled
Distributed: DeepSpeed ZeRO-3 across 3× H100 GPUs
Precision: BF16

Evaluation Results (Combined - Training Set)

Note: These results are from the sft_qwen3_4b_newrewards_backtoSFThints_butnoGRPOhints_afterdedupandidfix_multitemp_run1 (temperature 0.4) evaluation run (best P-C of this specific model). This model keeps the exact same configuration as the previous run (d450f3e02358050fb65ae426f5185a2e2b09d203), with the exception of re-introducing hints for SFT training:

Re-introduced hints for SFT training (GRPO remains hint-free).
Evaluation with concurrency of 32 for some runs.

Metric	Value
Accuracy	0.5599
F1 Score	0.6105
P-C (Pair-Correct)	26.196%
P-V (Pair-Vulnerable)	42.420%
P-B (Pair-Benign)	17.553%
P-R (Pair-Reversed)	13.829%

Full metrics JSON files are available in sft_qwen3_4b_eval_results. See the specific metrics from above here: sft_qwen3_4b_eval_results/sft_qwen3_4b_newrewards_backtoSFThints_butnoGRPOhints_afterdedupandidfix_multitemp_run1/metrics_temp0p4/metrics_combined_training.json.

Repository

Training code: SSL_research (commit: 2fe308cf27581b0dcb65c8aa51697d0f85db4575)

Downloads last month: -

Safetensors

Model size

196k params

Tensor type

BF16

Model tree for Andrefty/qwen3-4b-vuln-sft-research

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Finetuned

(1045)

this model