wangzhang/prometheus-datasets

---
license: agpl-3.0
task_categories:
- text-classification
language:
- en
tags:
- abliteration
- safety
- alignment
- prometheus
size_categories:
- 1K<n<10K
---

# Prometheus Evaluation Datasets

Curated prompt datasets used by [Prometheus](https://github.com/wuwangzhang1216/prometheus) for automated LLM abliteration optimization.

## Dataset Structure

| File | Count | Description |
|------|-------|-------------|
| `good_500/good_prompts_500.json` | 500 | Harmless prompts (iteration set) |
| `good_1000/good_prompts_1000.json` | 1000 | Harmless prompts (full set) |
| `harmful_500/harmful_prompts_500.json` | 500 | Harmful prompts (iteration set) |
| `harmful_1000/harmful_prompts_1000.json` | 1000 | Harmful prompts (full set) |

The 500-example sets are recommended for faster iteration (~2x speedup vs 1000 with no clear quality loss).

## Schema

Each entry contains:

| Field | Description |
|-------|-------------|
| `id` | Unique identifier |
| `prompt` | The prompt text |
| `source` | Generation source |
| `category` | Topic category |
| `language` | Language code |
| `policy_label` | Safety policy classification |
| `risk_level` | Risk level assessment |
| `topic_tag` | Topic tag |
| `sophistication` | Prompt sophistication level |
| `format` | Prompt format |
| `style_family` | Style classification |
| `design_goal` | Design intent |

## Usage

These datasets are used by Prometheus to compute refusal directions and evaluate abliteration effectiveness. See the [Prometheus README](https://github.com/wuwangzhang1216/prometheus) for details.

## License

[AGPL-3.0](https://www.gnu.org/licenses/agpl-3.0)