MILU

홈페이지 · 논문 1편

## 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 42 subjects, reflecting both general and culturally specific knowledge from India. ## Key Features - 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: 42 subjects covering a wide range of topics - Questions: ~85,000 multiple-choice questions - Cultural Relevance: Incorporates India-specific knowledge from regional and state-level examinations ## Dataset Statistics | Language | Total Questions | Translated Questions | Avg Words Per Question | |----------|-----------------|----------------------|------------------------| | Bengali | 7138 | 1601 | 15.72 | | Gujarati | 5327 | 2755 | 16.69 | | Hindi | 15450 | 115 | 20.63 | | Kannada | 6734 | 1522 | 12.83 | | Malayalam| 4670 | 1534 | 12.82 | | Marathi | 7424 | 1235 | 18.8 | | Odia | 5025 | 1452 | 15.63 | | Punjabi | 4363 | 2341 | 19.9 | | Tamil | 7059 | 1524 | 13.32 | | Telugu | 7847 | 1298 | 16.13 | | English | 14036 | - | 22.01 | | Total| 85073 | 15377 | 16.77 (avg) | ## Dataset Structure ### Test Set The test set consists of the MILU (Multi-task Indic Language Understanding) benchmark, which contains approximately 85,000 multiple-choice questions across 11 Indic languages. ### Validation Set The dataset includes a separate validation set of 9,157 samples that can be used for few-shot examples during evaluation. This validation set was created by sampling from each of the 42 subject tags, which were then condensed into 8 broader domains. This approach ensures a balanced representation across subjects and domains, allowing for consistent few-shot prompting across different models and experiments. ### 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 | | Math and Sciences | Astronomy and Astrophysics, Biology, Chemistry, Computer Science, Logical Reasoning, Mathematics, Physics | | Social Sciences | Anthropology, International Relations, Psychology, Public Administration, Social Welfare and Development, Sociology, Sports and Recreation | ## Usage Since this is a gated dataset, after your request for accessing the dataset is accepted, you can set your HuggingFace token: ``bash export HF_TOKEN=YOUR_TOKEN_HERE ` To load the MILU dataset for a Language: `python 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", 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 72% - 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 If you use MILU in your research, please cite our paper: ` @misc{verma2024milumultitaskindiclanguage, title={MILU: A Multi-task Indic Language Understanding Benchmark}, author={Sshubam Verma and Mohammed Safi Ur Rahman Khan and Vishwajeet Kumar and Rudra Murthy and Jaydeep Sen}, year={2024}, eprint={2411.02538}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2411.02538}, }`` ## License This dataset is released under the MIT License. ## Contact For any questions or feedback, please contact: - Sshubam Verma (sshubamverma@ai4bharat.org) - Mohammed Safi Ur Rahman Khan (safikhan@ai4bharat.org) - Rudra Murthy (rmurthyv@in.ibm.com) - Vishwajeet Kumar (vishk024@in.ibm.com) ## Links - GitHub Repository - Paper - Hugging Face Dataset