KHUjongseo/E2E_20K_test

E2E 20K Synthetic Direction

A large-scale synthetic video benchmark for evaluating VideoLLMs' directional reasoning.
20,000 videos of colored geometric shapes moving in four cardinal directions, with automatically generated MCQ annotations.

Directions

Direction	Description
up	Object moves toward the top of the frame
down	Object moves toward the bottom of the frame
left	Object moves toward the left of the frame
right	Object moves toward the right of the frame

Usage

from datasets import load_dataset

ds = load_dataset("YOUR_HF_ID/E2E_20K", split="train")

Data Format

{
  "id": 0,
  "video": "synthetic20k/train/down/circle_blue_001.mp4",
  "question": "Which direction does the blue circle move (camera perspective)?",
  "candidates": ["Left", "Down", "Right", "Up"],
  "answer": "B",
  "answer_text": "Down",
  "direction": "down"
}

Field	Type	Description
id	int	고유 인덱스
video	str	비디오 상대 경로
question	str	방향 질문 (question template 랜덤 샘플링)
candidates	list[str]	4개 보기 (순서 랜덤 셔플)
answer	str	정답 보기 레이블 (A/B/C/D)
answer_text	str	정답 텍스트 (Up / Down / Left / Right)
direction	str	GT 방향 소문자

Video Structure

synthetic20k/
├── train/
│   ├── up/
│   │   ├── circle_blue_001.mp4
│   │   ├── circle_red_001.mp4
│   │   └── ...
│   ├── down/
│   ├── left/
│   └── right/
└── val/
    ├── up/
    ├── down/
    ├── left/
    └── right/

Shapes & Colors

Shapes: circle, square, triangle
Colors: blue, red, green, yellow, white

Statistics

Split	Total	Up	Down	Left	Right
train	16,000	4,000	4,000	4,000	4,000
val	4,000	1,000	1,000	1,000	1,000
total	20,000	5,000	5,000	5,000	5,000

Question Templates

Each sample draws from a pool of question templates to reduce template bias:

• "Which direction does the {color} {shape} move (camera perspective)?"
• "As the clip progresses, which direction does the {color} {shape} move?"
• "Which direction does the {color} {shape} move within the frame?"
• "Which way does the {color} {shape} move in the image plane?"
• (and more...)

Related Datasets

Dataset	Description
E2E_real_object	Real-world objects, same direction task, smaller scale