SAINITHIN/DriveSafe_LLaMA_Adapter

DriveSafe LLaMA-Adapter (8B)

Hugging Face checkpoint repository for DriveSafe — mirrors the official OpenGVLab LLaMA-Adapter release weights used with the DriveSafe framework (ICRA 2026).

Links

Resource	URL
Model (this repo)	SAINITHIN/DriveSafe_LLaMA_Adapter
Source code	github.com/Sainithin-bit/DriveSafe
Base LLM (DriveSafe paper)	meta-llama/Meta-Llama-3.1-8B-Instruct
Related context pipeline	SAINITHIN/Exaplanation_Distillation
Upstream checkpoint	OpenGVLab/LLaMA-Adapter v1.0.0

Contents

File	Description
llama_adapter_len10_layer30_release.pth	OpenGVLab LLaMA-Adapter release checkpoint (PyTorch, Git LFS)
adapter_config.json	Architecture / training metadata
training_config.json	Paper hyperparameters (Sec. IV-B)

Model summary

DriveSafe finetunes Meta-Llama-3.1-8B-Instruct with LLaMA-Adapter on DRAMA caption–instruction–response triplets. Given a geometry-aware driving caption Cv, the adapter predicts:

• Binary risk label (Yes / No)
• Risk caption Cr
• Risk-related keywords K
• Risky object bounding box b

Safety suggestions are derived from keywords using the rule map in the DriveSafe GitHub repo (configs/safety_keyword_map.json).

Training details

• Framework: LLaMA-Adapter
• Base model: meta-llama/Meta-Llama-3.1-8B-Instruct
• Dataset: DRAMA (extended with safety-suggestion annotations)
• Batch size: 4
• Learning rate: 2e-5
• Weight decay: 0.01
• Epochs: 5
• GPU: NVIDIA A6000 (40 GB)
• Mixed precision + gradient checkpointing

Download

pip install huggingface_hub
huggingface-cli login

huggingface-cli download SAINITHIN/DriveSafe_LLaMA_Adapter \
  llama_adapter_len10_layer30_release.pth \
  adapter_config.json \
  --local-dir ./checkpoints/DriveSafe_LLaMA_Adapter

For Git LFS clone:

git lfs install
git clone https://huggingface.co/SAINITHIN/DriveSafe_LLaMA_Adapter
cd DriveSafe_LLaMA_Adapter
git lfs pull

Usage with DriveSafe repo

git clone https://github.com/Sainithin-bit/DriveSafe.git
cd DriveSafe

python risk_assessment/infer_risk.py \
  --caption-file outputs/captions/example.json \
  --adapter-repo SAINITHIN/DriveSafe_LLaMA_Adapter \
  --output outputs/risk

Requires Meta Llama 3.1 license acceptance on Hugging Face.

Citation

@inproceedings{artham2026drivesafe,
  title     = {{DriveSafe}: A Framework for Risk Detection and Safety Suggestions in Driving Scenarios},
  author    = {Artham, Sainithin and Dasgupta, Avijit and Gangisetty, Shankar and Jawahar, C. V.},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}

License

Specify before public release. Base Llama weights remain subject to the Meta Llama 3.1 Community License.