afterquery-in/indic-culture-bench

Indic Cultural Verification Benchmark v0.1

The Data Layer for Indian AI — Built in Bangalore for Bharat

AfterQuery builds data infrastructure that makes AI actually work for India — not translated, not approximated, but natively understood.

Website: https://www.afterquery.in Contact: hello@afterquery.in Status: In Development — 50-task pilot first

Overview

Existing Indic benchmarks test linguistic passability. This benchmark tests cultural verification — whether models can reason about Indian institutions, code-switching, and domain-specific contexts where errors have real-world consequences.

Target: 500 verifiable tasks across 3 languages

• Hindi: National reach, most widely spoken
• Kannada: Underserved Dravidian language, 50M+ speakers
• Hinglish: Real-world code-mixing, how urban India actually communicates

Domain Breakdown

Domain	Target Tasks	Tests
Government Scheme Eligibility	200	Institutional Knowledge, Cultural Context, Domain Accuracy
Hinglish Customer Support	200	Code-switching, Naturalness, Cultural Context
BFSI & Legal	100	Institutional Knowledge, Domain Accuracy, Safety

Examples: PM-KISAN eligibility logic, UPI disputes in Hinglish transcript → structured JSON, RBI circular QA, loan clause extraction.

Task Format

Each task is designed for programmatic verification:

{
  "prompt": "राजस्थान में किसान क्रेडिट कार्ड के लिए...",
  "gold_answer": "₹50,000 तक बिना गारंटी, 7% ब्याज...",
  "python_verifier": "def verify(response): ...",
  "language": "hindi",
  "domain": "government_schemes",
  "cultural_tags": ["regional_variation", "institutional"],
  "quality_scores": {
    "linguistic_fidelity": 0.95,
    "cultural_relevance": 0.92,
    "naturalness": 0.90,
    "diversity": 0.88,
    "domain_accuracy": 0.97,
    "safety": 0.99,
    "coverage": 0.85,
    "contamination": 1.0
  }
}

8 Quality Dimensions measured per task: Linguistic Fidelity, Cultural Relevance, Naturalness, Diversity, Domain Accuracy, Safety, Coverage, Contamination.

Why This Benchmark Is Different

1. Domain-specific cultural reasoning: Not "translate this sentence" but "Is this farmer eligible for PM-KISAN given these conditions?"
2. Programmatic verification: Every task includes Python verifier for reproducible, objective evaluation.
3. Code-switching as first-class input: Hinglish treated as valid communication, not noise.
4. Expert-validated answers: Gold answers validated by native speakers and domain experts.

Building on MILU, IndicGLUE, IndicIFEval — we focus specifically on verifiable cultural reasoning.

Distribution Plan

• 100 tasks: Public on HuggingFace (this repo) for community evaluation
• 400 tasks: Held-out for design partners and continuous improvement loop

Benchmark v0.1 is in development. Model evaluations and leaderboard will be published at https://www.afterquery.in/#leaderboard as evaluations run.

Intended Use

• Evaluation of LLMs for Indian cultural and institutional understanding
• Failure discovery for data improvement loops
• Research on code-switching and transliteration robustness

Not for training. Please do not train on public tasks if you intend to report scores.

License

MIT — Open for research and commercial evaluation.

Contact

Built by AfterQuery — Data Infrastructure for Indian AI Website: https://www.afterquery.in Email: hello@afterquery.in Location: Bangalore, India

© 2026 AfterQuery. Built in Bangalore for Bharat.