SurakshaEval

Existing LLM safety benchmarks focus primarily on English and Western contexts, often overlooking the linguistic diversity and culturally grounded safety risks present in other languages.

We introduce SurakshaEval, a novel safety benchmark composed of human-written prompts spanning real-world scenarios, explicitly designed for ten major Indian languages - Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil, and Telugu, along with English.

We benchmark state-of-the-art LLMs on SurakshaEval, establishing safety baselines and identifying recurring failure modes, including over-refusal and insufficient contextual awareness in regionally sensitive settings, particularly in native scripts.

These findings underscore the need for region-specific data and structured evaluations to build safe, ethical, and culturally aligned AI systems.

This release contains 1,484 English/Indic question pairs (2,968 prompts) across 10 regional Indic languages, alongside English. Each row contains a question pair: an English prompt (question_en) and its corresponding Indic-language prompt (question_indic) describing the same scenario in that regional context. Each pair therefore contributes two prompts, one in each language. See the published paper for the benchmark methodology and evaluation results.

Access requirements

SurakshaEval is a gated dataset. Users must provide their contact information (email address and Hugging Face username) and agree to the terms and conditions, if any, presented on Hugging Face in order to access this dataset.

Sign in to Hugging Face and submit an access request on the dataset page. Once access is granted, use an access token belonging to that same account for the loading and downloading of examples below.

Getting started

Install the libraries for loading and downloading the dataset:

pip install datasets huggingface_hub

After access is granted, set that account's Hugging Face access token before starting Python:

export HF_TOKEN="hf_your_access_token"

Replace the placeholder with your token. Hugging Face libraries automatically use HF_TOKEN for authentication.

The examples below use the Hugging Face repository SurakshaAI/SurakshaEval.

Load and access the dataset

from datasets import load_dataset

REPO_ID = 'SurakshaAI/SurakshaEval'

# Load every region or select an Indic-language configuration.
pairs = load_dataset(REPO_ID, "all", split="test")
regional_pairs = load_dataset(REPO_ID, 'assamese', split="test")

# Access the aligned questions and annotations.
example = pairs[0]
print(example["question_en"])
print(example["question_indic"])
print(example["language"], example["type_of_harm"])

english_questions = pairs["question_en"]
indic_questions = pairs["question_indic"]

Download the full dataset

Download all repository files, then load the question pairs from the JSONL files:

from pathlib import Path
from datasets import load_dataset
from huggingface_hub import snapshot_download

download_dir = snapshot_download(
    repo_id='SurakshaAI/SurakshaEval',
    repo_type="dataset",
    local_dir="surakshaeval",
)

files = sorted(str(path) for path in Path(download_dir).glob("data/*/test.jsonl"))
pairs = load_dataset("json", data_files={"test": files}, split="test")

Download one regional file

For example, download and load only the Assamese-context question pairs:

from datasets import load_dataset
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="SurakshaAI/SurakshaEval",
    repo_type="dataset",
    filename="data/assamese/test.jsonl",
    local_dir="surakshaeval",
)

regional_pairs = load_dataset("json", data_files={"test": path}, split="test")

Save a few example pairs

Save the first five Assamese-context English/Indic pairs to a UTF-8 JSONL file:

import json
from datasets import load_dataset

sample_pairs = load_dataset(
    "SurakshaAI/SurakshaEval", "assamese", split="test[:5]"
)

with open("surakshaeval_example_pairs.jsonl", "w", encoding="utf-8") as file:
    for pair in sample_pairs:
        file.write(json.dumps(pair, ensure_ascii=False) + "\n")

Replace assamese with another regional configuration, such as hindi, or use all to select from the combined dataset. For a single-file download, also change the region in filename.

See the Hugging Face guides for loading datasets and downloading dataset files.

Dataset structure

The default all configuration includes every region. Each region also has a configuration of its own. All configurations use a single test split because this release is intended for benchmark evaluation. They reference the same JSONL files, so the all configuration does not duplicate data on disk.

Region	Indic language	Question pairs
assamese	as	128
bengali	bn	149
gujarati	gu	110
hindi	hi	116
kannada	kn	163
malayalam	ml	126
marathi	mr	129
punjabi	pa	195
tamil	ta	157
telugu	te	211
Total		1,484

Column	Type	Meaning
id	string	Stable pair identifier: Indic ISO language code and source index.
index	int64	Original index within the region; not globally unique.
region	string	Source folder/context, such as hindi.
language	string	Indic ISO 639-1 language code, such as hi; English is always en.
question_en	string	Original English question, preserved verbatim.
question_indic	string	Original Indic-language question, preserved verbatim.
risk_area	string	Original risk-area annotation.
type_of_harm	string	Original harm category.
question_type	string	Original question-type annotation; blank values remain empty strings.

Intended use and limitations

Use the paired questions to compare safety behavior across English and the corresponding Indic language while keeping regional context fixed. The prompts include harmful or offensive content for evaluation of LLMs.

Keep each English/Indic pair together if creating other splits. Questions from different regions may overlap and are retained as distinct benchmark examples.

License

The SurakshaEval dataset is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). The Hugging Face license identifier is cc-by-nc-sa-4.0.

Sharing and adapting the dataset is permitted for noncommercial purposes with appropriate attribution, a license link, and an indication of changes. Shared adaptations must use the same license. See the full license terms.

Citation

If you use the SurakshaEval dataset, benchmark, or code, please cite our paper using the BibTeX entry below:

DOI: 10.1609/aies.v9i1.43143

@article{banerjee2026surakshaeval,
  title   = {{SurakshaEval}: {A}n {I}ndic {S}afety {B}enchmark for {M}ultilingual {LLMs}},
  author  = {Banerjee, Debopriyo and Kavitha, Kapil Rajesh and Borah, Angana and
             Han, Xudong and Wang, Yuxia and Krishnamurthy, Parameswari and
             Agarwal, Utkarsh and Kulkarni, Atharva and Lata, Swaran and
             Munot, Ayush and Sahnan, Dhruv and Singh, Aaryamonvikram and
             Nakov, Preslav and Choudhury, Monojit},
  journal = {Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society},
  volume  = {9},
  number  = {1},
  pages   = {325--336},
  year    = {2026},
  month   = oct,
  doi     = {10.1609/aies.v9i1.43143},
  url     = {https://doi.org/10.1609/aies.v9i1.43143}
}