sachaarbonel/qwen2.5-0.5b-jb2xl-merged

Jailbreak-Detector-2-XL → Qwen2.5-0.5B-Instruct (merge + ONNX)

This repo includes a small script to:

• Merge the LoRA madhurjindal/Jailbreak-Detector-2-XL into the base model Qwen/Qwen2.5-0.5B-Instruct and save the FULL merged weights.
• Export the merged model to ONNX (default task: text-generation-with-past).

Outputs are written under:

• models/qwen2.5-0.5b-jb2xl-merged/ (merged full HF model)
• onnx/qwen2.5-0.5b-jb2xl/ (ONNX export)

Note: You need internet access the first time to download from Hugging Face.

Quick Start (Docker)

docker build --no-cache -t jb2xl-merge .
docker run --rm -it \
  -e HF_HUB_ENABLE_HF_TRANSFER=1 \
  -v "$(pwd)":/workspace \
  jb2xl-merge \
  python scripts/merge_and_export.py \
    --base Qwen/Qwen2.5-0.5B-Instruct \
    --adapter madhurjindal/Jailbreak-Detector-2-XL \
    --out models/qwen2.5-0.5b-jb2xl-merged \
    --onnx-out onnx/qwen2.5-0.5b-jb2xl \
  --task text-generation-with-past

If you have a private HF token, pass it at runtime (Transformers picks it up automatically):

docker run --rm -it \
  -e HUGGINGFACE_HUB_TOKEN=hf_xxx \
  -v "$(pwd)":/workspace \
  jb2xl-merge \
  python scripts/merge_and_export.py

If you see ModuleNotFoundError: No module named 'optimum.exporters.onnx', rebuild the image with --no-cache as shown above, or install the extras inside the container:

docker run --rm -it -v "$(pwd)":/workspace jb2xl-merge \
  pip install "optimum[exporters,onnxruntime]>=1.23.3"

Local Python (optional)

If you prefer not to use Docker, install Python 3.10+ and run:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python scripts/merge_and_export.py

Arguments

--base        Base model (default: Qwen/Qwen2.5-0.5B-Instruct)
--adapter     LoRA adapter (default: madhurjindal/Jailbreak-Detector-2-XL)
--out         Output dir for merged model (default: models/qwen2.5-0.5b-jb2xl-merged)
--onnx-out    Output dir for ONNX (default: onnx/qwen2.5-0.5b-jb2xl)
--task        ONNX task: text-generation | text-generation-with-past (default)
--opset       ONNX opset (optional)
--dtype       Torch dtype for merging load (default: float32)
--no-onnx     Skip ONNX export

Notes

• The merged model is saved with safetensors format and includes the tokenizer.
• ONNX export uses Optimum and will generate the appropriate graph(s) for the task you choose.
• For very large ONNX files (>2GB), external data format is handled automatically by Optimum.