license: mit
configs:
- config_name: default
data_files:- split: train
path: data/train-* - split: validation
path: data/validation-* - split: test
path: data/test-*
dataset_info:
features: - name: conversations
list:- name: from
dtype: string - name: value
dtype: string
splits:
- name: from
- name: train
num_bytes: 3031451.3566189012
num_examples: 9379 - name: validation
num_bytes: 378810.2132378028
num_examples: 1172 - name: test
num_bytes: 379133.4301432958
num_examples: 1173
download_size: 2021820
dataset_size: 3789395.0
- split: train
SafetyBreak Dataset
A dataset for binary + multiclass safety classification.
Overall Stats
- Total examples: 11,724
- Labels:
- Safe: 4,500 (38.38%)
- Unsafe: 7,224 (61.62%)
- Unsafe categories:
- Harmful Instructions: 2,000 (27.69%)
- Self Harm: 2,000 (27.69%)
- Toxic Content: 2,000 (27.69%)
- Malicious Code: 1,224 (16.94%)
Splits
| Split | Examples | Safe (%) | Unsafe (%) |
|---|---|---|---|
| Train | 9379 | 38.09 | 61.91 |
| Validation | 1172 | 39.59 | 60.41 |
| Test | 1173 | 39.56 | 60.44 |
Intended Use
Designed for fine-tuning models on AI guardrail systems:
- Binary classification: Safe vs Unsafe
- Multiclass classification: Threat type (toxic_content, self_harm, harmful_instructions, malicious_code)