README.md

metadata

license: apache-2.0
task_categories:
  - question-answering
  - text-classification
  - summarization
language:
  - en
  - hi
pretty_name: IndicConBench
size_categories:
  - 1K-10K
tags:
  - legal
  - legal-nlp
  - constitutional-reasoning
  - indian-constitution
  - multilingual
  - constitutional-qa
  - zero-leakage
  - vikhram-labs
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: validation.jsonl
      - split: test
        path: test.jsonl
      - split: gpqa
        path: gpqa.jsonl

IndicConBench: A Multilingual Constitutional Reasoning Benchmark for India

IndicConBench is a publication-grade, research-quality multilingual constitutional reasoning benchmark for India. It is specifically designed to evaluate the factual accuracy, reasoning capacity, structural retrieval capability, and legal intelligence of Large Language Models (LLMs) on the Constitution of India in both English and Hindi.

Developed by Vikhram S and published under Vikhram Labs, this dataset serves as a rigorous scientific instrument for the multilingual legal AI domain.

Dataset Summary

• Dataset Name: IndicConBench
• Creator/Author: Vikhram S (Vikhram Labs)
• Organization Page: Vikhram Labs on Hugging Face
• Repository Link: https://huggingface.co/datasets/vikhram-labs/IndicConBench
• Languages: English (en), Hindi (hi)
• Total Articles Covered: 465
• Total Parallel Examples: 8408
• Estimated Total Tokens: 1050651
• Splits:
  • Train: 6708 examples
  • Validation: 832 examples
  • Test: 848 examples
  • GPQA (IndicConBench-GPQA): 20 examples (Expert-authored Multiple Choice)

---

Task Families

IndicConBench contains 9 diverse task families designed to evaluate distinct aspects of legal AI:

1. Constitutional QA (qa): Factual question-answering pairs testing constitutional facts.
2. Article Retrieval (retrieval): Locating specific constitutional articles given their functional guarantees.
3. Constitutional Reasoning (reasoning): Analyzing complex real-world legal scenarios to identify constitutional violations and underlying principles.
4. Summarization (summarization): Evaluating generation capabilities at various granularities (One-line, Short summaries).
5. Entailment (entailment): Natural Language Inference (NLI) pairs testing logical entailment, contradiction, and neutral claims.
6. Multi-hop Reasoning (multihop): Integrating knowledge across multiple distinct articles to answer legal questions.
7. Principle Classification (classification): Categorizing articles into key parts (Fundamental Rights, DPSP, Parliament, etc.).
8. Article Linking (linking): Identifying functional relations (supports, extends, limits, references, related_to) between article pairs.
9. IndicConBench-GPQA (gpqa): Graduate-level expert-authored constitutional reasoning questions requiring multi-step reasoning, precedent interaction, and distractor elimination.

---

Dataset Splits & Leakage Prevention

To ensure strict compliance with NLP evaluation standards, a rigorous article-level partition is performed (80% Train, 10% Validation, 10% Test):

• All questions, scenarios, and entailment hypotheses generated from a given article reside in the same split.
• No data leakage occurs between the training set and validation/testing sets.
• The IndicConBench-GPQA split is expert-authored and held out as a strict diagnostic set for measuring advanced reasoning limits.

---

Quick Start (Usage)

You can easily load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("vikhram-labs/IndicConBench")

# Access splits
train_data = dataset["train"]
gpqa_data = dataset["gpqa"]

# Preview an example
print(gpqa_data[0])

---

Scientific Evaluation Protocol

To benchmark models against the golden test split, we provide a structured Python evaluation script (evaluate.py) which computes Exact Match (EM), macro F1-score, and ROUGE-L metrics across tasks and languages.

Run the evaluation:

python evaluate.py --predictions <model_predictions.jsonl> --gold dataset/test.jsonl

---

Citation & Attribution

If you use this benchmark in your research, please cite the work as follows:

@misc{vikhramlabs2026indicconbench,
  author = {Vikhram S},
  title = {IndicConBench: A Multilingual Constitutional Reasoning Benchmark for India},
  year = {2026},
  publisher = {Vikhram Labs},
  journal = {Hugging Face Datasets},
  howpublished = {\url{https://huggingface.co/datasets/vikhram-labs/IndicConBench}}
}

---

License: Apache 2.0 Disclaimer: This dataset is created for evaluating language models on constitutional reasoning and does not constitute formal legal counsel.