README.md

metadata

language:
  - en
license: apache-2.0
pretty_name: ViTaB-A
task_categories:
  - question-answering
tags:
  - table-question-answering
configs:
  - config_name: hitab
    data_files:
      - split: train
        path: hitab/train.jsonl
      - split: validation
        path: hitab/validation.jsonl
      - split: test
        path: hitab/test.jsonl
  - config_name: fetaqa
    data_files:
      - split: train
        path: fetaqa/train.jsonl
      - split: validation
        path: fetaqa/validation.jsonl
      - split: test
        path: fetaqa/test.jsonl

ViTaB-A Dataset

A normalized table question answering dataset for the ViTaB-A research project.

Configs

• hitab: Derived from HiTab (10,670 samples)
• fetaqa: Derived from FeTaQA (10,330 samples)

Usage

from datasets import load_dataset

hitab = load_dataset("raoanmol/ViTaB-A", "hitab")
fetaqa = load_dataset("raoanmol/ViTaB-A", "fetaqa")

Schema

Each sample contains:

Field	Type	Description
id	string	Unique identifier (e.g. vitaba_000001_hitab)
split	string	Dataset split (train/validation/test)
question	string	Natural language question about the table
answer	list or string	Answer (list for HiTab, string for FeTaQA)
citation	list[str]	Excel-style cell references (e.g. ["=E7"])
table_json	dict	Simplified table with keys: title (string), header (list of header rows), rows (list of data rows)
table_md	string	Markdown representation of the table with Excel-style row/column labels
table_images	dict	Table images as base64 PNGs. Keys: arial, times_new_roman, red, blue, green. Unrendered variants are empty strings.
source	string	Source dataset and split (e.g. hitab_train)
source_id	string	Original ID from source dataset