MILU-1.5: A Multi-task Indic Language Understanding Benchmark (cleaned)

Overview

MILU (Multi-task Indic Language Understanding Benchmark) is a comprehensive evaluation dataset designed to assess the performance of Large Language Models (LLMs) across 11 Indic languages. It spans 8 domains and 41 subjects, reflecting both general and culturally specific knowledge from India.

Key Features

• 11 Indian Languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, and English
• Domains: 8 diverse domains including Arts & Humanities, Social Sciences, STEM, and more
• Subjects: 41 subjects covering a wide range of topics
• Questions: 84,690 multiple-choice questions (3,851 defective items removed from the original 88,541)
• Cultural Relevance: Incorporates India-specific knowledge from regional and state-level examinations

What is different in MILU-1.5

MILU-1.5 is a cleaned release of ai4bharat/MILU. The benchmark, its schema, its 11 languages, its 8 domains and 41 subjects, its splits and its option order are all unchanged. No question was added, no answer key was altered, and no option was reordered or renumbered. Only two things happened: defective items were removed, and a small number were repaired in place.

Every one of the 88,541 released items was audited by an LLM judge. The findings were grouped into error types, and eight of them were acted on:

Removed without further review — unrepairable by construction:

• Missing figure or passage — the stem points at a figure, chart, map, table or reading passage that the row never contained, so the item cannot be answered from its own text.
• Duplicate or identical options — two of the four options are the same.

Reviewed item by item by gemini-3.7-flash, one request per row carrying all of that row's findings at once. For each finding the model decided whether it was a false positive, a repair forced by the surrounding text, or a reconstruction that would need guesswork; any single unrepairable finding removed the row:

• Incomplete question or numbered statements
• Malformed question or options
• Translation or spelling defect
• Mixed-language content — correct, purposeful English (technical terms, proper nouns, abbreviations) was left alone; only garbled bilingual concatenations and broken fragments were repaired
• Underspecified or unverifiable question
• Other invalid-item defect

Of 5,788 rows reviewed this way, 709 were repaired, 1,863 were confirmed to be false positives and left byte-identical, and 3,216 were removed. Typical repairs: 15\n1\n3\n% restored to 15 1/3%; a collapsed 1.496 × 1011 restored to 1.496 × 10¹¹; एफएलएफए corrected to फीफा; an untranslated English option rendered into the item's own language.

Known limitations — what MILU-1.5 does not fix

These error types were deliberately left untouched and are still present:

• wrong answer key — the largest remaining class by far
• wrong domain, subject, language or is_translated metadata
• ambiguous items, or items with several defensible answers
• items where none of the four options is correct
• factual or reasoning errors in the stem

A wrong answer key never caused a row to be removed. MILU-1.5 should be read as repaired on eight structural defect types, not as a fully corrected benchmark.

Repair and removal decisions are model judgements over an LLM judge's findings; neither pass was verified by hand at scale.

MILU-1.5 dataset statistics

Counts are of rows actually present in the released Parquet files.

Language	Test (v1)	Test (1.5)	Validation (v1)	Validation (1.5)	Removed	Repaired	Removed %
Bengali	6,637	6,343	812	734	372	46	4.99%
English	13,535	13,432	813	807	109	160	0.76%
Gujarati	4,826	4,498	813	731	410	39	7.27%
Hindi	14,831	14,681	812	744	218	112	1.39%
Kannada	6,234	5,967	812	716	363	43	5.15%
Malayalam	4,321	3,935	813	729	470	38	9.15%
Marathi	6,924	6,652	810	728	354	63	4.58%
Odia	4,525	4,170	812	737	430	40	8.06%
Punjabi	4,099	3,768	811	743	399	27	8.13%
Tamil	6,372	6,098	813	726	361	65	5.02%
Telugu	7,304	7,031	812	720	365	76	4.50%
Total	79,608	76,575	8,933	8,115	3,851	709	4.35%

Dataset statistics of the original MILU release

Language	Total Questions	Translated Questions	Avg Words Per Question
Bengali	6638	1601	15.12
Gujarati	4827	2755	16.12
Hindi	14837	115	20.61
Kannada	6234	1522	12.42
Malayalam	4321	3354	12.39
Marathi	6924	1235	18.76
Odia	4525	3100	14.96
Punjabi	4099	3411	19.26
Tamil	6372	1524	13.14
Telugu	7304	1298	15.71
English	13536	-	22.07
Total	79617	19915	16.41 (avg)

Dataset Structure

Test Set

The test set consists of the MILU (Multi-task Indic Language Understanding) benchmark, which contains approximately 80,000 multiple-choice questions across 11 Indic languages.

Validation Set

The dataset includes a separate validation set of 8,933 samples that can be used for few-shot examples during evaluation. This validation set was created by sampling questions from each of the 41 subjects.

Subjects spanning MILU

Domain	Subjects
Arts & Humanities	Architecture and Design, Arts and Culture, Education, History, Language Studies, Literature and Linguistics, Media and Communication, Music and Performing Arts, Religion and Spirituality
Business Studies	Business and Management, Economics, Finance and Investment
Engineering & Tech	Energy and Power, Engineering, Information Technology, Materials Science, Technology and Innovation, Transportation and Logistics
Environmental Sciences	Agriculture, Earth Sciences, Environmental Science, Geography
Health & Medicine	Food Science, Health and Medicine
Law & Governance	Defense and Security, Ethics and Human Rights, Law and Ethics, Politics and Governance
Science	Astronomy and Astrophysics, Biology, Chemistry, Computer Science, Logical Reasoning, Physics
Social Sciences	Anthropology, International Relations, Psychology, Public Administration, Social Welfare and Development, Sociology, Sports and Recreation

Usage

Since this is a private dataset, set your HuggingFace token before loading it:

export HF_TOKEN=YOUR_TOKEN_HERE

To load the MILU dataset for a Language:

from datasets import load_dataset

language = 'Hindi'

# Use 'test' split for evaluation & 'validation' split for few-shot
split = 'test'

language_data = load_dataset("ai4bharat/MILU-1.5", data_dir=language, split=split, token=True)

print(language_data[0])

Evaluation

We evaluated 45 different LLMs on MILU, including:

• Closed proprietary models (e.g., GPT-4o, Gemini-1.5)
• Open-source multilingual models
• Language-specific fine-tuned models

Key findings:

• GPT-4o achieved the highest average accuracy at 74%
• Open multilingual models outperformed language-specific fine-tuned models
• Models performed better in high-resource languages compared to low-resource ones
• Performance was lower in culturally relevant areas (e.g., Arts & Humanities) compared to general fields like STEM

For detailed results and analysis, please refer to our paper.

Citation

MILU-1.5 is a cleaned release of MILU and has no separate paper. If you use it, please cite the MILU paper:

@inproceedings{verma-etal-2025-milu,
    title = "{MILU}: A Multi-task {I}ndic Language Understanding Benchmark",
    author = "Verma, Sshubam  and
      Khan, Mohammed Safi Ur Rahman  and
      Kumar, Vishwajeet  and
      Murthy, Rudra  and
      Sen, Jaydeep",
    editor = "Chiruzzo, Luis  and
      Ritter, Alan  and
      Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.507/",
    doi = "10.18653/v1/2025.naacl-long.507",
    pages = "10076--10132",
    ISBN = "979-8-89176-189-6",
    abstract = "Evaluating Large Language Models (LLMs) in low-resource and linguistically diverse languages remains a significant challenge in NLP, particularly for languages using non-Latin scripts like those spoken in India. Existing benchmarks predominantly focus on English, leaving substantial gaps in assessing LLM capabilities in these languages. We introduce MILU, a Multi-task Indic Language Understanding Benchmark, a comprehensive evaluation benchmark designed to address this gap. MILU spans 8 domains and 41 subjects across 11 Indic languages, reflecting general and culturally specific knowledge. With an India-centric design, incorporates material from regional and state-level examinations, covering topics such as local history, arts, festivals, and laws, alongside standard subjects like science and mathematics. We evaluate over 42 LLMs, and find that current LLMs struggle with MILU, with GPT-4o achieving the highest average accuracy at 74 percent. Open multilingual models outperform language-specific fine-tuned models, which perform only slightly better than random baselines. Models also perform better in high resource languages as compared to low resource ones. Domain-wise analysis indicates that models perform poorly in culturally relevant areas like Arts and Humanities, Law and Governance compared to general fields like STEM. To the best of our knowledge, MILU is the first of its kind benchmark focused on Indic languages, serving as a crucial step towards comprehensive cultural evaluation. All code, benchmarks, and artifacts are publicly available to foster open research."
}

License

This dataset is released under the CC BY 4.0.

Contact

For any questions or feedback, please contact:

• Sshubam Verma (verma.sshubam@gmail.com)
• Mohammed Safi Ur Rahman Khan (safikhan@ai4bharat.org)
• Rudra Murthy (rmurthyv@in.ibm.com)
• Vishwajeet Kumar (vishk024@in.ibm.com)

Links

• Original dataset 🤗
• GitHub Repository 💻
• Paper 📄