IndicKLAR

IndicKLAR (Indic Knowledge and Language-consistency Assessment Resource) is a factual question-answering dataset for evaluating cross-lingual consistency of LLMs across 18 Indic languages and English, plus romanized/code-mixed variants.

Paper preprint: arXiv:2605.29637 · Code: IndicKLAR_EMNLP_2026

Summary

• 78,570 samples — 2,619 samples × 30 language/script variants
• 20 relations (e.g. capital, place_of_birth, official_language, occupation)
• 19 native-script languages (18 Indic languages + English) + 11 romanized "code-mixed" (-en suffixed) variants for transliteration/code-mixing experiments
• Used to compute the Cross-Lingual Consistency (CLC) score — whether an LLM answers the same fact correctly regardless of language, script, or code-mixing strategy (formula and evaluation harness in the GitHub repo)

Structure

data/<lang>/<relation>.json   # e.g. data/hin/place_of_birth.json

Each config in the viewer dropdown corresponds to one language/script directory (hin, hin-en, ben, ...), since native-script directories carry extra English-gloss fields that romanized/English ones don't — keeping them as separate configs avoids merging mismatched schemas.

from datasets import load_dataset
ds = load_dataset("debajyotimaz/IndicKLAR", "hin")
print(ds["train"][0])

Fields

Each .json file is {"name": <relation>, "prompt_templates": [...], "samples": [...]}. Per-sample fields:

Field	Description
index	Sample index, stable across all 30 language variants for a given relation
subject / object	Subject entity and ground-truth answer, in the source script
object_candidates	Comma-separated multiple-choice distractors + correct answer
subject_en / object_en	English gloss (native-script configs only)
relation	Relation name, denormalized onto every sample

Relations (20)

applies_to_jurisdiction, capital, capital_of, continent, country_of_citizenship, developer, field_of_work, headquarters_location, instrument, language_of_work_or_name, languages_spoken, location_of_formation, manufacturer, native_language, occupation, official_language, owned_by, place_of_birth, place_of_death, religion

Languages (30 configs)

Hindi, Bengali, Assamese, Gujarati, Telugu, Malayalam, Marathi, Odia, Punjabi, Tamil, Sanskrit — each with a -en romanized/code-mixed sibling config — plus Kannada, Urdu, Sindhi, Dogri, Konkani, Maithili, Nepali (native-script only), and English.

Usage

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download("debajyotimaz/IndicKLAR", "data/hin/place_of_birth.json",
                        repo_type="dataset")
print(json.load(open(path, encoding="utf-8"))["samples"][0])

Dataset Creation

Facts spanning geography, biography, culture, and organizations were rendered into natural-language question templates for each of the 18 Indic languages plus English. Romanized code-mixed (-en) variants were derived from the native-script versions. Multiple-choice distractors were sampled per relation to support closed-form evaluation alongside open-ended generation.

Intended Uses

Evaluating factual accuracy and cross-lingual consistency of multilingual LLMs; studying prompting strategy (direct, code-mixed, transliterated, translated, latent reasoning) effects on Indic-language factual QA; benchmarking low-resource language understanding.

License

CC BY 4.0. Evaluation code is released separately under MIT at IndicKLAR_EMNLP_2026.

Citation

Preprint citation (will be updated once the EMNLP 2026 proceedings version is published):

@article{mazumder2026evaluating,
  title={Evaluating Cross-lingual Knowledge Consistency in Code-Mixed vis-a-vis Indian Languages using IndicKLAR},
  author={Mazumder, Debajyoti and Pathak, Divyansh and Kodali, Prashant and Joshi, Aditya and Agarwal, Akshay and Patro, Jasabanta},
  journal={arXiv preprint arXiv:2605.29637},
  year={2026}
}