license: mit
task_categories:
- text-classification
- text-generation
language: - en
tags: - alignment
- over-refusal
- safety
pretty_name: Alignment 4-Type
size_categories: - 1K<n<10K
Alignment 4-Type Dataset
8-type alignment data for over-refusal research. 4 prompt categories × 2 response behaviors.
Data Design
| category_id | Type | Prompt Category | Response Behavior | Train | Val |
|---|---|---|---|---|---|
| 0 | unsafe_refusal | unsafe (harmful) | refusal | 200 | 50 |
| 1 | unsafe_helpful | unsafe (harmful) | helpful | 0 | 50 |
| 2 | safe_refusal | safe (harmless) | refusal | 0 | 50 |
| 3 | safe_helpful | safe (harmless) | helpful | 200 | 50 |
| 4 | unknown_refusal | unknown (ambiguous) | uncertainty | 200 | 50 |
| 5 | unknown_helpful | unknown (ambiguous) | confident | 0 | 50 |
| 6 | known_refusal | known (factual) | refusal | 0 | 50 |
| 7 | known_helpful | known (factual) | confident | 200 | 50 |
- Train: 800 total (4 types × 200)
- Val: 400 total (8 types × 50)
- Val pairs share the same prompts within each category (refusal and helpful responses to the same prompt)
- All data quality-filtered via DeepSeek LLM-as-judge (type_match ≥ 4, coherent, no leakage, relevant)
Files
train.parquet— 800 training samplesval.parquet— 400 validation samples
Fields
| Field | Type | Description |
|---|---|---|
prompt |
string | user prompt |
response |
string | model response |
category |
string | one of the 8 type names (e.g. unsafe_refusal) |
category_id |
int8 | numeric id 0-7, see table above |
Generation
- Responses generated by Qwen2.5-7B (local vLLM)
- Refusal/uncertainty: direct prompting with varied system prompts
- Helpful/confident: prefilled assistant starts to bypass safety filters
- Quality filtering: DeepSeek API judge (type_match ≥ 4 + quality_ok = true)
Usage
from datasets import Dataset
train = Dataset.from_parquet("train.parquet")
val = Dataset.from_parquet("val.parquet")