---
language:
- si
license: cc-by-4.0
task_categories:
- text-classification
- question-answering
tags:
- sinhala
- low-resource
- safety
- alignment
- refusal
- responsible-ai
- cultural
pretty_name: Culturally Aware Sinhala Dataset
size_categories:
- n<1K
---

# Culturally Aware Sinhala Dataset

A hand-curated Sinhala (සිංහල) diagnostic dataset of **safety-critical, culturally
grounded** prompts paired with the safe, responsible response a model is expected
to give. It is designed as a **sanity-check benchmark** for the safety and
responsibility of LLM responses in Sinhala, a low-resource language with strong
cultural grounding.

Modern LLMs are increasingly used as primary information sources. Safety and
alignment are well studied for high-resource languages such as English, but their
effectiveness in low-resource languages is largely unexplored. Multilingual models
are trained on highly imbalanced corpora, so safety behaviour that holds in English
may not generalise to Sinhala. This dataset probes exactly that gap.

## Responsible, not merely harmless

This dataset extends the usual "harmless yet non-evasive" criterion with a
**risk-awareness (safety-oriented)** requirement: a good response must not just
refuse or avoid harm, it must **actively discourage harmful actions** and provide
**appropriate, culturally grounded guidance** in high-risk contexts (e.g. local
Sri Lankan helpline numbers in self-harm refusals).

## Dataset structure

| Column            | Description                                                     |
|-------------------|-----------------------------------------------------------------|
| `question`        | User prompt in Sinhala (safety-critical / culturally grounded). |
| `category`        | Safety/topic label for the prompt (see below).                  |
| `expected_answer` | The safe, responsible, expected Sinhala response.               |

- **Rows:** 102
- **Language:** Sinhala (`si`)
- **Splits:** single `train` split

### Categories

| category            | count |
|---------------------|-------|
| criminal            | 24    |
| medical             | 20    |
| hate                | 10    |
| weapons             | 8     |
| cultural            | 8     |
| violence            | 7     |
| illegal_substance   | 6     |
| sexual              | 6     |
| help                | 5     |
| fraud               | 3     |
| profanity           | 3     |
| suicide             | 2     |

Categories span self-harm, medical advice, culturally sensitive topics, illegal
activities, weapons, sexual content, and abusive language.

## Intended use

- **Diagnostic benchmark** for the safety of Sinhala LLM responses.
- Evaluating whether safety mechanisms generalise from high-resource languages
  to a low-resource, culturally grounded language.
- Fine-tuning / alignment data for culturally aware, responsible Sinhala responses.

## Limitations

- Small size (102 examples); a sanity check, not a comprehensive benchmark.
- Expected answers reflect one set of safety norms and the Sri Lankan context;
  they may not generalise to other locales.
- Contains references to sensitive topics (self-harm, violence, illegal acts) by
  design.

## Paper

Introduced in our paper on the safety of LLM responses in Sinhala.

## License

Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): free to
use, share and adapt, including commercially, with attribution.