Add model card and metadata

README.md

@@ -0,0 +1,58 @@
+---
+library_name: transformers
+pipeline_tag: video-text-to-text
+tags:
+- traffic-safety
+- video-understanding
+- autonomous-driving
+---
+
+# CrashChat: A Multimodal Large Language Model for Multitask Traffic Crash Video Analysis
+
+CrashChat is a Multimodal Large Language Model (MLLM) specifically designed for traffic crash video analysis. Built upon VideoLLaMA3, it provides a unified framework for tasks such as crash recognition, temporal grounding, and high-level understanding (description and reasoning) across diverse driving scenarios.
+
+- **Paper:** [CrashChat: A Multimodal Large Language Model for Multitask Traffic Crash Video Analysis](https://arxiv.org/abs/2512.18878)
+- **Repository:** [https://github.com/Liangkd/CrashChat](https://github.com/Liangkd/CrashChat)
+
+## Introduction
+
+CrashChat addresses the challenges of automating crash video analysis by incorporating three key contributions:
+1. **Multitask Learning Strategy**: Employs task decoupling and grouping to maximize joint learning benefits while mitigating negative transfer between linguistic-centric and perception-centric tasks.
+2. **Unified Framework**: Capable of performing crash recognition, temporal grounding, and reasoning within a single model.
+3. **Domain-Specific Fine-tuning**: Trained on a comprehensive instruction-tuning dataset comprising 18,385 videos and over 96,000 video-QA pairs.
+
+Numerical experiments demonstrate that CrashChat achieves state-of-the-art performance, reaching near-perfect accuracy in crash recognition and significant improvements in both crash localization and causal reasoning compared to general-purpose MLLMs.
+
+## Model Zoo
+
+The project includes several fine-tuned checkpoints based on different training strategies:
+
+| Checkpoint | MLLM backbone | Training Strategy - Task |
+|----------|-------------|-------------|
+| VideoLLaMA-3 | VideoLLaMA-3 7B | VideoLLaMA3 - baseline |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - crash recognition |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - crash description |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - causal reasoning |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - prevention reasoning |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - pre-crash localization |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Independent monotask models - crash localization |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Homogeneous multitask models - linguistic-centric tasks |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Homogeneous multitask models - perception-centric tasks |
+| CrashChat-7B-Finetuned | VideoLLaMA-3 7B | Heterogeneous multitask models - all tasks |
+
+## Citation
+
+If you find CrashChat useful for your research, please cite:
+
+```bibtex
+@article{liang2025crashchat,
+  title   = {CrashChat: A Multimodal Large Language Model for Multitask Traffic Crash Video Analysis},
+  author  = {Liang, Kaidi and Li, Ke and Hu, Xianbiao and Qin, Ruwen},
+  journal = {arXiv preprint arXiv:2512.18878},
+  year    = {2025},
+  archivePrefix = {arXiv},
+  eprint  = {2512.18878},
+  primaryClass = {cs.CV},
+  url     = {https://arxiv.org/abs/2512.18878}
+}
+```