README.md

---
language:
- en
pretty_name: OpenSafetyMini
size_categories:
- 1K<n<10K
- 10K<n<100K
task_categories:
- text-generation
- text-classification
tags:
- safety
- llm-safety
- quantization
- human-evaluation
- llm-as-a-judge
configs:
- config_name: prompts
  default: true
  data_files:
  - split: test
    path: data/OpenSafetyMini.csv
- config_name: human_annotations
  data_files:
  - split: test
    path: data/OpenSafetyMini_responses_annotated.csv
---

# OpenSafetyMini

OpenSafetyMini is a hard, open-ended safety benchmark for LLMs. 

It has **1,067 human-curated English prompts** and **21,328 human safety labels** for model responses. 

The dataset was released with the paper [*Investigating the Impact of Quantization Methods on the Safety and Reliability of Large Language Models*](https://ieeexplore.ieee.org/iel8/6287639/6514899/11563799.pdf).

- **Paper:** https://ieeexplore.ieee.org/iel8/6287639/6514899/11563799.pdf
- **Code & prompts:** https://github.com/On-Point-RND/OpenSafetyMini-Investigating-the-Impact-of-Quantization-Methods-on-the-Safety-and-Reliability-of-LLM

## How it was built

The prompts come from the English part of [XSAFETY](https://github.com/Jarviswang94/Multilingual_safety_benchmark), which has many vague or ambiguous questions. They were filtered in two stages:

1. **Deflection score.** GPT-4o rated each prompt 10 times on how likely a model is to refuse it on ethical grounds (0–100). Prompts with an average score below 50 were removed.
2. **Manual review.** The remaining prompts were checked by hand, and low-quality ones were dropped.

The result is 1,067 prompts that challenge modern models and separate safe from unsafe models more clearly than the original set.

## Configurations

### `prompts` (default): 1,067 rows

| Column | Description |
|---|---|
| `prompt_id` | Prompt ID (shared with `human_annotations`) |
| `lang` | Language (always `en`) |
| `scenario` | Safety category (see below) |
| `prompt` | The prompt text |

Prompts per scenario:

| Scenario | # |
|---|---|
| Unsafe_Instruction_Topic | 189 |
| Goal_Hijacking | 181 |
| Crimes_And_Illegal_Activities | 172 |
| Unfairness_And_Discrimination | 149 |
| Role_Play_Instruction | 145 |
| Privacy_And_Property | 124 |
| Mental_Health | 73 |
| Reverse_Exposure | 34 |

[`data/opensafetymini_safety_criteria.json`](data/opensafetymini_safety_criteria.json) describes what counts as an unsafe response in each scenario.

### `human_annotations`: 21,328 rows

These are responses from four versions of [Llama 3.1 8B Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct), each labeled separately by five annotators. That gives 4,268 unique question–answer pairs × 5 annotators.

| Column | Description |
|---|---|
| `prompt_id` | Prompt ID (matches `prompts`) |
| `method` | Model version: `FP 16` (original), `QUIK 4bit`, `QUIP# 2bit`, `Abliterated` (refusal direction removed) |
| `annotator_id` | Annotator (`0`–`4`) |
| `label` | `safe`, `unsafe`, `ambiguous`, or `error` |
| `prompt` | The prompt text |
| `response` | Model response |

Label meanings:
- **safe / unsafe**: whether the response is safe, judged by the scenario criteria.
- **ambiguous**: the annotator could not decide if it is safe or unsafe.
- **error**: gibberish or repeated text that cannot be judged.

The annotators were five linguists with experience in LLM safety evaluation, paid four times the minimum wage. Agreement between them (Krippendorff's α) is 0.623. The model versions were shown in random order.

**Building a single reference label.** The paper takes a majority vote across the five annotators. Ties are broken toward the more cautious label: `unsafe` > `ambiguous` > `safe` > `error`.

## Usage

```python
from datasets import load_dataset

prompts = load_dataset("On-Point-Rnd/OpenSafetyMini", "prompts", split="test")
annotations = load_dataset("On-Point-Rnd/OpenSafetyMini", "human_annotations", split="test")
```

The `prompts/` folder has the evaluation prompt templates used in the paper, filled with Python's `str.format`:
- [`opensafetymini_llm_judge_evaluation_prompt_template.md`](prompts/opensafetymini_llm_judge_evaluation_prompt_template.md): LLM-as-a-Judge template. With Gemma 2 27B as the judge, it agrees with the human reference label 92% of the time.
- [`safetybench_5shot_prompt_template.md`](prompts/safetybench_5shot_prompt_template.md): SafetyBench 5-shot template.
- [`hotpotqa_generation_prompt_template.md`](prompts/hotpotqa_generation_prompt_template.md): HotPotQA generation template.

## Intended use and limitations

The dataset is meant for **evaluating** LLM safety, especially how compression and fine-tuning affect it. Do not use it to train models to produce harmful content. It is English-only, and its labels reflect the judgment of a small group of annotators.

## Citation

```bibtex
@article{kharinaev2025investigating,
  title   = {Investigating the Impact of Quantization Methods on the Safety and Reliability of Large Language Models},
  author  = {Kharinaev, Artyom and Moskvoretskii, Viktor and Shvetsov, Egor and Studenikina, Kseniia and Mikhail, Bykov and Burnaev, Evgeny},
  journal = {arXiv preprint arXiv:2502.15799},
  year    = {2025}
}
```