README.md

---
language:
- en
license: mit
task_categories:
- video-text-to-text
extra_gated_fields:
  Name: text
  Company/Organization: text
  Country: text
  E-Mail: text
modalities:
- Video
- Text
configs:
- config_name: event_sequence
  data_files: json/event_sequence.json
- config_name: moving_direction
  data_files: json/moving_direction.json
- config_name: reversible_dynamics
  data_files: json/reversible_dynamics.json
---

# DyBench

[**Project Page**](https://ddz16.github.io/crpo.github.io/) | [**Paper**](https://huggingface.co/papers/2605.21988) | [**GitHub**](https://github.com/ddz16/CRPO)

DyBench is a paired counterfactual video benchmark introduced in the paper "[Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning](https://huggingface.co/papers/2605.21988)". 

The benchmark is designed to evaluate the **spatiotemporal sensitivity** of Video Large Language Models (Video LLMs). It addresses the issue of models relying on "shortcuts" (such as single-frame cues or language priors) rather than tracking actual video dynamics. DyBench utilizes a strict pair-accuracy metric that requires a model to correctly answer questions for both original and counterfactual versions of a video.

### Dataset Details
DyBench consists of **3,014 videos** covering three primary categories of spatiotemporal dynamics:
- **Reversible Dynamics**: Evaluating if models understand physical processes that can be temporally reversed.
- **Moving Direction**: Tracking the spatial trajectory and direction of motion.
- **Event Sequence**: Understanding the temporal order in which events occur.

### Data Structure
The dataset is organized into three configurations corresponding to the tasks above:
- `event_sequence`
- `moving_direction`
- `reversible_dynamics`

Each configuration contains JSON files mapping videos to their respective questions and ground-truth answers.