README.md

metadata

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