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.
CrashQwen-8B access is gated to document noncommercial research use and attribution/citation agreement.
Log in or Sign Up to review the conditions and access this model content.
CrashQwen-8B
CrashQwen-8B is a domain-adapted version of Qwen3-VL-8B-Instruct for structured multi-view post-crash analysis.
The model was adapted on Crash-VQA using LoRA applied to the language-model attention and feed-forward projections while keeping the vision encoder frozen.
Released checkpoint
This repository contains the final adapter used for the principal Crash-VQA paper results.
- Base model:
Qwen/Qwen3-VL-8B-Instruct - Training run:
rosy-sweep-9 - W&B run ID:
cz6hwyti - Selected checkpoint: step 2250
- LoRA rank: 16
- LoRA alpha: 32
- LoRA dropout: 0
- Metadata policy: metadata for Delta-V and AIS2+ tasks only
- Training seed: 42
On the complete natural Crash-VQA test split, the adapted model achieves 67.3% micro accuracy compared with 45.4% for the original Qwen3-VL-8B-Instruct zero-shot model.
Files
adapter_model.safetensors-- final LoRA adapteradapter_config.json-- PEFT adapter configurationcrashvqa_training_config.json-- training configuration associated with the released adapter
The base Qwen3-VL-8B-Instruct weights are not redistributed in this repository.
Dataset
Crash-VQA: https://huggingface.co/datasets/oValach/Crash-VQA
License and intended use
This release is intended for noncommercial research use. See the repository license and access conditions before using the checkpoint.
The model is a research artifact and should not be used as the sole basis for medical, legal, insurance, accident-liability, or operational vehicle-safety decisions.
- Downloads last month
- -
Model tree for oValach/CrashQwen-8B
Base model
Qwen/Qwen3-VL-8B-Instruct