Tacha is a dataset derived from TAT-QA, containing financial documents with tables, designed to evaluate how well chunking algorithms handle structured tabular data mixed with narrative text.

Dataset Description

• Documents: 349 financial documents with tables
• Questions: 2,065 question-answer pairs
• Domain: Financial Tables
• Source: TAT-QA dataset

Key Challenges

This dataset tests chunking algorithms on:

• Tabular data structures
• Numerical reasoning across rows/columns
• Table headers and cell relationships
• Mixed table and text content
• Financial calculations and comparisons
• Cross-references between tables and narrative

Dataset Structure

Corpus Config

Questions Config

Usage

from datasets import load_dataset

# Load corpus
corpus = load_dataset("chonkie-ai/tacha", "corpus", split="train")

# Load questions
questions = load_dataset("chonkie-ai/tacha", "questions", split="train")

Part of MTCB

Tacha is part of the Massive Text Chunking Benchmark (MTCB), a comprehensive benchmark for evaluating RAG chunking strategies.

Citation

If you use this dataset, please also cite the original TAT-QA paper.

License

CC-BY-4.0