README.md

metadata

language:
  - en
pretty_name: TallyBench
tags:
  - counting
  - visual reasoning
  - benchmark
license: cc-by-4.0
task_categories:
  - visual-question-answering
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_name
      dtype: string
    - name: vlm_question
      dtype: string
    - name: gt_answer
      dtype: int32
    - name: categories
      dtype: string
    - name: image_type
      dtype: string
  splits:
    - name: train
      num_bytes: 1062256505
      num_examples: 2000
  download_size: 1000274834
  dataset_size: 1062256505

TallyBench

TallyBench is a dataset of 2,000 counting images with QA pairs.
It was constructed to evaluate quantity reasoning in vision–language models (VLMs).
TallyBench also serves as an auxiliary dataset for CompareBench.

Related Datasets

• CompareBench
• HistCaps

Code

👉 CompareBench on GitHub