metadata
license: mit
language:
- en
task_categories:
- question-answering
- text-generation
tags:
- scientific-reasoning
- video-understanding
- physics
- chemistry
- benchmark
VideoScienceBench
A benchmark for evaluating video understanding and scientific reasoning in vision-language models. Each example pairs a textual description of an experiment (what is shown) with the correct scientific explanation (expected phenomenon).
Dataset Summary
| Attribute | Value |
|---|---|
| Examples | 160 |
| Domains | Physics, Chemistry |
| Format | JSONL (prompt + expected phenomenon + vid) |
Data Creation Pipeline
Each researcher selects two or more scientific concepts and references relevant educational materials or videos to design a prompt. Prompts undergo peer and model review, followed by model-based quality checking, before being finalized for dataset inclusion.
Dataset Structure
Each line is a JSON object with:
| Field | Type | Description |
|---|---|---|
| keywords | list[str] | Relevant scientific concepts |
| field | str | Scientific discipline (e.g., Physics) |
| prompt | str | Textual description of what is shown in the video/experiment |
| expected phenomenon | str | The correct scientific explanation |
| vid | str | Video identifier |
Example
{
"keywords": ["Buoyancy", "Gas Laws", "Pressure"],
"field": "Physics",
"prompt": "A sealed plastic bottle is filled with water containing a floating eyedropper with an air bubble inside. A person squeezes the sides of the bottle.",
"expected phenomenon": "The eyedropper immediately sinks when the bottle is squeezed, then rises again when released, as increased pressure compresses the air bubble, reducing buoyancy.",
"vid": "98"
}
Usage
from datasets import load_dataset
dataset = load_dataset("lmgame/VideoScienceBench", data_files="data_filtered.jsonl")
# Access the data (single split)
data = dataset["train"]
License
MIT