RoadTones-VL-CoT

This model accompanies the paper RoadTones: Tone Controllable Text Generation from Road Event Videos

Model Summary

RoadTones-VL-CoT is an open-source large multimodal model with superior tone-controlled road video captioning capabilities. Built on the foundation of Qwen3-VL-8B-Instruct, it has been finetuned on RoadTones-51k dataset with the Chain-of-Thought (CoT) intermediate drafts as well for better interpretability. Evaluated on the RoadTones-Eval Metrics, Its performance is on par with the best benchmarked model (Gemini-2.5-pro) while displaying superior Tone Adherance, thereby demonstrating the RoadTones-51K dataset's capability in improving the tone-controllability for text generation in VideoLLMs.

For further details, please refer to the following resources:

• 🪐 Project Page: https://roadtones.github.io
• 📦 Dataset: https://huggingface.co/datasets/siddhi-lipare/RoadTones
• 💻 Code: https://github.com/roadtones/roadtones
• 📰 Paper: https://arxiv.org/abs/2605.21411

Citation

@misc{parikh2026roadtonestonecontrollabletext,
      title={RoadTones: Tone Controllable Text Generation from Road Event Videos}, 
      author={Chirag Parikh and Siddhi Pravin Lipare and Ravi Kiran Sarvadevabhatla},
      year={2026},
      eprint={2605.21411},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.21411}, 
}