README.md

metadata

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 | Paper | GitHub

DyBench is a paired counterfactual video benchmark introduced in the paper "Learning Spatiotemporal Sensitivity in Video LLMs via Counterfactual Reinforcement Learning".

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.