JoeyCCC/VidNum-1.4K

VidNum: Video-Grounded Numerical Reasoning Benchmark

VidNum is a manually curated benchmark for evaluating video-grounded numerical reasoning in vision-language models. The current validated release contains 1,167 multiple-choice questions derived from 947 source videos. Each question is grounded in a video segment and asks models to identify, count, track, compare, or compose quantities from visual evidence.

This repository has been updated to the current benchmark version. The old single-file release is superseded by two language-specific files:

• data/vidnum_en.jsonl: English questions, options, answers, and full labels.
• data/vidnum_zh.jsonl: Chinese questions, options, answers, and full labels.

Both files contain the same 1,167 question IDs. The only difference is the question/option language and localized label names.

Important Video Access Note

Each example includes an original video_source_link and timestamp. These fields are the recommended way to identify the source material.

The repository may still contain a videos/ directory with clip files from the earlier release. These files are retained for compatibility with existing evaluation scripts, but the current metadata release is designed so users can trace every question back to the original online video and timestamp. Only the question IDs listed in data/vidnum_en.jsonl and data/vidnum_zh.jsonl belong to the current benchmark. Some legacy clip files may remain in videos/ for backward compatibility and should not be treated as additional current benchmark examples.

For compatibility, the clip filename is always determined directly from the question ID:

id = 1     -> videos/QID_1.mp4
id = 858   -> videos/QID_858.mp4
id = 1167  -> videos/QID_1167.mp4

Dataset Files

File	Language	Rows	Description
data/vidnum_en.jsonl	English	1,167	English questions and options with full annotations.
data/vidnum_zh.jsonl	Chinese	1,167	Chinese questions and options with full annotations.
dataset_summary.json	Mixed	1	Release statistics and label counts.

Each JSONL line is one question. The id values are stable and shared across the English and Chinese files.

Field Schema

Core fields:

Field	Type	Description
id	integer	Stable question ID. This also determines the compatible clip path videos/QID_{id}.mp4.
source_platform	string	Source platform inferred from the original collection path, e.g. bilibili or douyin.
source_video_id	string	Identifier of the original online source video. This is not the clip filename.
video_source_link	string	Original online video URL.
timestamp	string	Temporal segment for the question in MM:SS-MM:SS or equivalent format.
video_file	string	Compatible clip filename, always QID_{id}.mp4.
video_path	string	Compatible clip path, always videos/QID_{id}.mp4.
language	string	en or zh.
category	string	Broad video category.
sub_category	string	Fine-grained video category.
question	string	Question text in the file language.
option_A	string	Option A in the file language.
option_B	string	Option B in the file language.
option_C	string	Option C in the file language.
option_D	string	Option D in the file language.
answer	string	Gold answer label, one of A, B, C, or D.
answer_text	string	Gold answer option text in the file language.

Diagnostic annotation fields:

Field	Description
evidence_target	Primary evidence type to quantify: object, action, or event. In the Chinese file this is localized; the code is in evidence_target_code.
count_scope	Counting scope or structural counting requirement. In the Chinese file this is localized; the code is in count_scope_code.
reasoning_type	Required reasoning operation, e.g. none, calculation, comparison, logic, estimation, or mixed. In the Chinese file this is localized; the code is in reasoning_type_code.
scene_structure	Video structure label, e.g. edited-with-cuts or continuous-shot. In the Chinese file this is localized; the code is in scene_structure_code.
fine_grained_count_type	Fine-grained counting structure label. In the Chinese file this is localized; the English code is in fine_grained_count_type_code.
task_group	Main VidNum diagnostic task group. In the Chinese file this is localized; the English code is in task_group_code.

Main Diagnostic Task Groups

VidNum uses a computation-oriented taxonomy. These labels describe the quantitative demand imposed by each question; they should not be interpreted as an automatic easy-to-hard ordering.

English Label	Chinese Label	Meaning
Direct and Distinct Enumeration	直接与去重枚举	The target set is directly specified; the model must enumerate visible instances, entities, or categories.
Conditioned and Structured Enumeration	条件与结构化枚举	The target set must be constructed from attributes, relations, temporal intervals, event boundaries, or aggregation rules before counting.
Compositional Quantitative Reasoning	组合式定量推理	The answer requires an explicit operation over one or more grounded quantities, such as calculation, comparison, ordering, logical inference, or estimation.

Fine-Grained Counting Types

English Label	Chinese Label
Direct Homogeneous Enumeration	直接同质枚举
Distinct Entity or Category Enumeration	不同实体或类别枚举
Attribute- or Relation-Conditioned Enumeration	属性或关系条件枚举
Multi-Instance or Event Aggregation	多实例或事件聚合
Compositional Quantitative Reasoning	组合式定量推理

Label Distribution

Current release statistics:

Dimension	Counts
Questions	1,167
Source videos	947
Evidence targets	object: 543, action: 332, event: 292
Task groups	DDE: 458, CSE: 407, CQR: 302

Loading Examples

The benchmark name is VidNum. The Hugging Face repository slug is currently kept as JoeyCCC/VidNum-1.4K for compatibility with earlier releases.

Using datasets:

from datasets import load_dataset

en = load_dataset("JoeyCCC/VidNum-1.4K", "english", split="test")
zh = load_dataset("JoeyCCC/VidNum-1.4K", "chinese", split="test")

example = en[0]
print(example["question"])
print(example["video_source_link"], example["timestamp"])
print(example["video_path"])  # videos/QID_1.mp4

Using plain Python:

import json

with open("data/vidnum_en.jsonl", encoding="utf-8") as f:
    for line in f:
        item = json.loads(line)
        assert item["video_file"] == f"QID_{item['id']}.mp4"

Evaluation Recommendation

For multiple-choice evaluation, present the model with the video segment and the four answer options, and require the final answer to be one of A, B, C, or D. Report the language, frame sampling strategy, prompt format, and whether the compatible local clip files or original source videos were used.

License and Use

The VidNum annotations and question files are released under CC BY 4.0. Evaluation code is released under the license used in the project repository.

The original online videos are owned by their respective creators or platforms. VidNum does not claim ownership of those videos. The current metadata release provides source links and timestamps for research reproducibility.

Citation

If you use VidNum, please cite:

@misc{cui2026vidnum14k,
  title         = {VidNum-1.4K: A Comprehensive Benchmark for Video-based Numerical Reasoning},
  author        = {Cui, Shaoyang and Meng, Lingbei and Luo, Yaodi and He, Peize},
  year          = {2026},
  eprint        = {2604.03701},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2604.03701},
  url           = {https://arxiv.org/abs/2604.03701}
}

Paper: https://arxiv.org/abs/2604.03701