JoeyCCC/VidNum-1.4K

---
pretty_name: VidNum-1.4K
language:
- en
- zh
license: other
task_categories:
- question-answering
task_ids:
- visual-question-answering
- multiple-choice-qa
size_categories:
- 1K<n<10K
annotations_creators:
- expert-generated
source_datasets:
- original
tags:
- video
- counting
- numerical reasoning
configs:
- config_name: default
data_files:
- split: train
path: VidNum1_4K.jsonl
---

# VidNum-1.4K

VidNum-1.4K is a multiple-choice video-based numerical reasoning QA benchmark.
Each sample contains a source video link, a timestamp clip range, a question, four choices, and one gold answer.

## Dataset Summary

- Annotation file: `VidNum1_4K.jsonl`
- Number of QA samples: `1379`
- Video files found under `videos/`: `1379` mp4 files

## Task

- Primary task: video counting question answering
- Format: 4-way multiple choice (`A/B/C/D`)
- Reasoning labels: `none`, `calculation`, `comparison`, `logic`, `mixed`, `estimation`

## Data Fields

The JSONL contains the following keys:

- `ID`: integer sample id
- `Video Source Link`: source URL of the original video
- `Video_path`: corresponding video file.
- `Category`: top-level category
- `Sub-Category`: fine-grained category
- `Timestamp`: clip range in `mm:ss-mm:ss`
- `question`: English question text
- `option_A`: option A
- `option_B`: option B
- `option_C`: option C
- `option_D`: option D
- `Answer`: correct option in `{A, B, C, D}`
- `Level`: difficulty level in `{1, 2, 3}`
- `Reasoning_Type`: reasoning tag
- `Count_Scope`: counting scope tag