qcri-ai/HCTQA

HCT-QA: Human-Centric Tables Question Answering

HCT-QA is a benchmark dataset designed to evaluate large language models (LLMs) on question answering over complex, human-centric tables (HCTs). These tables often appear in documents such as research papers, reports, and webpages and present significant challenges for traditional table QA due to their non-standard layouts and compositional structure.

The dataset includes:

• 2,188 real-world tables with 9,835 human-annotated QA pairs
• 4,679 synthetic tables with 67,500 programmatically generated QA pairs
• Logical and structural metadata for each table and question

📄 Paper: [Title TBD]
The associated paper is currently under review and will be linked here once published.

---

How to load in Python (as pandas DataFrames):

from datasets import load_dataset
import pandas as pd

dataset = load_dataset("qcri-ai/HCTQA")
    
# Convert each split to a pandas DataFrame
train_df = pd.DataFrame(dataset['train'])
val_df = pd.DataFrame(dataset['validation'])
test_df = pd.DataFrame(dataset['test'])

---

📊 Dataset Splits

Config	Split	# Examples (Placeholder)
RealWorld	Train	7,500
RealWorld	Test	2,335
Synthetic	Train	55,000
Synthetic	Test	12,500

---

🏆 Leaderboard

Model Name	FT (Finetuned)	Recall	Precision
Model-A	True	0.81	0.78
Model-B	False	0.64	0.61
Model-C	True	0.72	0.69

📌 If you're evaluating on this dataset, open a pull request to update the leaderboard.

---

Dataset Structure

Each entry in the dataset is a dictionary with the following structure:

Sample Entry

{
  "table_id": "arxiv--1--1118",
  "dataset_type": "arxiv",
  "table_data": {
    "table_as_csv": ",0,1,2\n0,Domain,Average Text Length,Aspects Identified\n1,Journalism,50,44\n...",
    "table_as_html": "<table><tr><th>Domain</th><th>Average Text Length</th>...",
    "table_as_markdown": "| Domain | Average Text Length | Aspects Identified |...",
    "table_image_local_path_within_github_repo": "tables/images/arxiv--1--1118.jpg",
    "table_image_url": "https://hcsdtables.qcri.org/datasets/all_images/arxiv_1_1118.jpg",
    "table_properties_metadata": {
      "Standard Relational Table": true,
      "Row Nesting": false,
      "Column Aggregation": false
    }
  },
  "questions": [
    {
      "question_id": "arxiv--1--1118--M0",
      "question": "Report the Domain and the Average Text Length where the Aspects Identified equals 72",
      "question_template_for_synthetic_only": "Report [column_1] and [column_2] where [column_3] equals [value]",
      "question_properties_metadata": {
        "Row Filter": true,
        "Aggregation": false,
        "Returned Columns": true
      },
      "answer": "{Psychology | 86} || {Linguistics | 90}",
      "prompt": "<system>...</system><user>...</user>",
      "prompt_without_system": "<user>...</user>"
    }
  ]
}

Ground Truth Format

Explain the GT format here
Example: {value1 | value2} || {value3 | value4}

Table Properties

For details on table and question properties please see our paper