Skip to content

Repository files navigation

BhashaBench

CC BY 4.0

Overview

BhashaBench is a series of India-specific, multi-task benchmarks designed to evaluate Large Language Models (LLMs) across multiple Indic languages, diverse domains, and varied task types. Built from the ground up with authentic Indian contexts, it goes beyond factual knowledge to capture cultural nuances, domain-specific expertise, and rural values. The suite includes datasets spanning agriculture, governance, education, and more offering a mix of question answering, classification, reasoning, translation, and other tasks to ensure robust, multilingual, and context-aware performance assessment.

This repository contains code for evaluating language models on the BBK benchmark using the lm-eval-harness framework.

Domain Specific Benchmark

  1. BhashaBench-Krishi (BBK): It is the first large-scale, authentic benchmark designed to rigorously evaluate AI models on Indian agricultural knowledge. Tailored for India’s diverse agro-ecological zones, crops, languages, and farming practices, BBK draws from 55+ official government agricultural exams to assess models' ability to provide precise, region-aware, policy-relevant, and actionable agricultural advice. Hugging Face

    Key Features

    • Languages: English and Hindi (with plans for more Indic languages)
    • Exams: 55+ unique agricultural government and institutional exams across India
    • Domains: 25+ agricultural and allied science domains, spanning over 270 topics
    • Questions: 15,405 rigorously validated, exam-based questions
    • Difficulty Levels: Easy (6,754), Medium (6,941), Hard (1,710)
    • Question Types: Multiple Choice, Assertion-Reasoning, Match the Column, Rearrange the Sequence, Fill in the Blanks
    • Focus: Practical, context-rich, region-specific agricultural knowledge essential for Indian farmers

    Dataset Structure

    Test Set

    The test set consists of the BBK benchmark, which contains approximately 15,000 multiple-choice questions across 2 Indic languages.

    Dataset Statistics

    Metric Count
    Total Questions 15,405
    English Questions 12,648
    Hindi Questions 2,757
    Subject Domains 25+
    Government Exams Covered 55+

    Subjects spanning BBK

    Subject Domain Count
    Agri-Environmental & Allied Disciplines 176
    Agricultural Biotechnology 524
    Agricultural Chemistry & Biochemistry 281
    Agricultural Economics & Policy 627
    Agricultural Engineering & Technology 244
    Agricultural Extension Education 774
    Agricultural Microbiology 111
    Agriculture Communication 254
    Agriculture Information Technology 190
    Agronomy 5078
    Animal Sciences 148
    Crop Sciences 549
    Dairy & Poultry Science 89
    Entomology 696
    Fisheries and Aquaculture 34
    General Knowledge & Reasoning 661
    Genetics and Plant Breeding 389
    Horticulture 2070
    Natural Resource Management 193
    Nematology 184
    Plant Pathology 397
    Plant Sciences & Physiology 129
    Seed Science and Technology 202
    Soil Science 1357
    Veterinary Sciences 48
  2. BhashaBench-Finance (BBF): It is the first comprehensive benchmark designed to evaluate AI models on Indian financial knowledge and practices. Tailored for India's diverse financial ecosystem, regulatory framework, and economic landscape, BBF draws from 25+ official government and institutional financial exams to assess models' ability to provide accurate, policy-compliant, and contextually relevant financial advice and analysis. Hugging Face

    Key Features

    • Languages: English and Hindi (with plans for more Indic languages)
    • Exams: 25+ unique financial government and institutional exams across India
    • Domains: 30+ financial and allied domains, spanning comprehensive financial knowledge
    • Questions: 19,433 rigorously validated, exam-based questions
    • Difficulty Levels: Easy (7,111), Medium (9,348), Hard (2,974)
    • Question Types: Multiple Choice, Assertion-Reasoning, Match the Column, Rearrange the Sequence, Fill in the Blanks, Reading Comprehension, Essay
    • Focus: Practical, regulation-aware, India-specific financial knowledge essential for financial professionals and consumers

    Dataset Structure

    Test Set

    The test set consists of the BBF benchmark, which contains approximately 19,000 questions across 2 Indic languages.

    Dataset Statistics

    Metric Count
    Total Questions 19,433
    English Questions 13,451
    Hindi Questions 5,982
    Subject Domains 30+
    Government Exams Covered 25+

    Subjects spanning BBF

    Subject Domain Count
    Problem Solving 5,686
    Mathematics for Finance 4,845
    Banking Services 1,171
    Governance & Policy 1,064
    Language & Communication 946
    Corporate Finance & Investment 910
    Commerce 863
    Accounting 773
    General Knowledge 539
    Information Technology Finance 490
    Economics & Development Studies 274
    Rural Economics 261
    Environmental Finance 168
    Taxation & Regulatory Compliance 155
    Interdisciplinary Finance 153
    Data & Analytics in Finance 127
    History, Sociology & Cultural Studies of Finance 127
    Finance Education 118
    Healthcare Economics 114
    Science and Technology in Finance 101
    International Finance & Trade 83
    Business Management 83
    Energy, Infrastructure & Finance 82
    Behavioral Finance 67
    Financial Markets 47
    Sports, Media & Finance Linkages 45
    Marketing Finance 42
    Insurance & Risk Management 42
    Legal Finance 34
    Financial Technology 23
  3. BhashaBench-Legal (BBL) is the first large-scale, authentic benchmark designed to rigorously evaluate AI models on Indian legal knowledge.
    Tailored for India’s diverse jurisdictional contexts, laws, exams, and legal practices, BBL draws from 50+ official government and institutional law exams to assess models' ability to provide precise, context-aware, policy-relevant, and actionable legal advice. Hugging Face

    Key Features

    • Languages: English and Hindi (with plans for more Indic languages)
    • Exams: 50+ unique legal government and institutional exams across India
    • Domains: 20+ legal and allied disciplines, spanning over 200 specialized topics
    • Questions: 24,365 rigorously validated, exam-based questions
    • Difficulty Levels: Easy (8,200), Medium (12,150), Hard (4,015)
    • Question Types: Multiple Choice, Assertion-Reasoning, Match the Column, Rearrange the Sequence, Fill in the Blanks
    • Focus: Practical, context-rich, jurisdiction-specific legal knowledge essential for Indian legal practice

    Dataset Structure

    Test Set

    The test set consists of the BBL benchmark, which contains approximately 24,365 multiple-choice questions across 2 Indic languages (English and Hindi).
    Support for more Indic languages will be added in upcoming versions.

    Dataset Statistics

    Metric Count
    Total Questions 24,365
    English Questions 17,047
    Hindi Questions 7,318
    Subject Domains 20+
    Government Exams Covered 50+

    Subjects spanning BBL

    Subject Domain Count
    Civil Litigation & Procedure 7126
    Constitutional & Administrative Law 3609
    Criminal Law & Justice 2769
    Corporate & Commercial Law 2700
    General Academic Subjects 1756
    Legal Theory & Jurisprudence 1421
    Family & Personal Law 991
    International & Comparative Law 962
    Legal Skills & Communication 816
    Real Estate & Property Law 629
    Environmental & Energy Law 430
    Interdisciplinary Studies 363
    Tax & Revenue Law 231
    Employment & Labour Law 175
    Technology & Cyber Law 123
    Intellectual Property Law 91
    Consumer & Competition Law 75
    Media & Entertainment Law 54
    Healthcare & Medical Law 25
    Human Rights & Social Justice 19
  4. BhashaBench-Ayur (BBA) is the first comprehensive, large-scale benchmark designed to rigorously evaluate AI models on Ayurvedic knowledge and practice.
    Tailored for India's traditional medicine system, BBA draws from authentic Ayurvedic texts, examinations, and clinical practices to assess models' ability to provide accurate, context-aware, and clinically relevant guidance in Ayurveda. Hugging Face

    Key Features

    • Languages: English and Hindi (with plans for more Indic languages)
    • Sources: Authentic Ayurvedic examinations, classical texts, and modern Ayurvedic education standards
    • Domains: 15+ specialized Ayurvedic disciplines covering comprehensive traditional medicine knowledge
    • Questions: 14,963 rigorously validated, examination-based questions
    • Difficulty Levels: Easy (7,944), Medium (6,314), Hard (705)
    • Question Types: Multiple Choice Questions, Fill in the Blanks, Match the Column, Assertion-Reasoning
    • Focus: Practical, clinically-relevant, traditional medicine knowledge essential for Ayurvedic practice

    Dataset Structure

    Test Set

    The test set consists of the BBA benchmark, which contains approximately 14,963 questions across 2 Indic languages (English and Hindi).
    Support for more Indic languages will be added in upcoming versions.

    Dataset Statistics

    Metric Count
    Total Questions 14,963
    English Questions 9,348
    Hindi Questions 5,615
    Subject Domains 15+
    Classical Texts Covered Multiple

    Question Type Distribution

    Question Type Count
    Multiple Choice (MCQ) 14,717
    Fill in the Blanks 178
    Match the Column 41
    Assertion or Reasoning 27

    Difficulty Level Distribution

    Difficulty Level Count
    Easy 7,944
    Medium 6,314
    Hard 705

    Subjects spanning BBA

    Subject Domain Count
    Kayachikitsa (General Medicine & Internal Medicine) 3,134
    Dravyaguna & Bhaishajya (Pharmacology & Therapeutics) 2,972
    Samhita & Siddhanta (Fundamental Principles) 1,541
    Sharir (Anatomy & Physiology) 1,346
    Panchakarma & Rasayana (Detoxification & Rejuvenation) 1,308
    Stri Roga & Prasuti Tantra (Gynecology & Obstetrics) 847
    Shalakya Tantra (ENT, Ophthalmology & Dentistry) 734
    Kaumarbhritya & Pediatrics (Child Health) 714
    Agad Tantra & Forensic Medicine (Toxicology) 587
    Shalya Tantra (Surgery) 526
    Swasthavritta & Public Health (Preventive Medicine) 453
    Research & Statistics 210
    Ayurvedic Literature & History 204
    Yoga & Psychology 188
    Administration, AYUSH & Miscellaneous 119
    Roga Vigyana (Diagnostics & Pathology) 80

Usage

Prerequisites
  • Python 3.7+
  • lm-eval-harness library
  • HuggingFace Transformers
  • vLLM (optional, for faster inference)
  1. Clone this repository:
git clone --depth 1 https://github.com/BharatGen-IITB-TIH/BhashaBench.git
cd BhashaBench
pip install -e .
  1. Request access to the HuggingFace 🤗 dataset here.

  2. Set up your environment variables:

export HF_HOME=/path/to/HF_CACHE/if/needed
export HF_TOKEN=YOUR_HUGGINGFACE_TOKEN

The following languages are supported for BBK, BBF, BBL and BBA:

  • English
  • Hindi

HuggingFace Evaluation

For HuggingFace models, you may use the following sample command:

lm_eval --model hf \
    --model_args 'pretrained=google/gemma-2-27b-it,temperature=0.0,top_p=1.0,parallelize=True' \
    --tasks bbk,bbf,bbl,bba \
    --batch_size auto:40 \  
    --log_samples \
    --output_path $EVAL_OUTPUT_PATH \
    --max_batch_size 64 \
    --apply_chat_template

vLLM Evaluation

For vLLM-compatible models, you may use the following sample command:

lm_eval --model vllm \
    --model_args 'pretrained=meta-llama/Llama-3.2-3B,tensor_parallel_size=$N_GPUS' \
    --gen_kwargs 'temperature=0.0,top_p=1.0' \
    --tasks bbk,bbf,bbl,bba \
    --batch_size auto \
    --log_samples \
    --output_path $EVAL_OUTPUT_PATH

Single Language Evaluation

To evaluate your Model on a specific language, modify the --tasks parameter:

--tasks bbk_English
--tasks bbf_Hindi

Replace English with the available language (e.g., Hindi, etc.).

Evaluation Tips & Observations

  1. Make sure to use --apply_chat_template for Instruction-fine-tuned models, to format the prompt correctly.
  2. vLLM generally works better with Llama models, while Gemma models work better with HuggingFace.
  3. If vLLM encounters out-of-memory errors, try reducing max_gpu_utilization else switch to HuggingFace.
  4. For HuggingFace, use --batch_size=auto:<n_batch_resize_tries> to re-select the batch size multiple times.
  5. When using vLLM, pass generation kwargs in the --gen_kwargs flag. For HuggingFace, include them in model_args.

License

This dataset is released under the CC BY 4.0.

Links

About

No description, website, or topics provided.

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages