Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is granted manually for research on AI safety only. This dataset contains harmful, offensive, and sexual text.
Log in or Sign Up to review the conditions and access this dataset content.
Qwen3.5-SafeProbe Training Data
Dataset Summary
Qwen3.5-SafeProbe Training Data is a bilingual (Vietnamese and English) dataset designed for training and evaluating AI safety classification models.
The dataset contains up to 90K high-quality samples, collected from publicly available datasets and custom-curated sources. Selected samples have been manually or programmatically reviewed and relabeled to improve annotation correctness, consistency, and alignment with the intended safety policies.
The dataset is developed to support two safety models:
- Qwen3.5-SafeProbe — A general-purpose safety classification model designed for bilingual content moderation and safety-related detection.
- Qwen3.5-SafeProbe-VI — A Vietnamese-focused variant customized for Vietnamese safety policies, linguistic nuances, and locally relevant content moderation requirements.
The primary goal is to provide reliable training data for building lightweight, specialized safety models that can serve as guardrails in LLM applications and AI agent systems.
Intended Use
This dataset is intended for research and development in AI safety, particularly:
- Training text-based safety classifiers.
- Building input and output guardrails for LLM applications.
- Detecting potentially harmful or policy-violating content.
- Identifying prompt injection and jailbreak attempts.
- Developing Vietnamese-specific content moderation systems.
- Studying cross-lingual safety classification and policy alignment.
- Evaluating the robustness and consistency of safety detection models.
Dataset Characteristics
| Attribute | Description |
|---|---|
| Dataset name | Qwen3.5-SafeProbe Training Data |
| Maximum size | Up to 90K samples |
| Languages | Vietnamese, English |
| Data type | Text |
| Primary task | Text classification |
| Domain | AI safety and content moderation |
| Data sources | Public datasets and custom-curated data |
| Annotation | Reviewed and relabeled samples |
| License | CC-BY-SA-4.0 |
Supported Safety Domains
The dataset is designed to cover safety-related content, including:
- Harmful Content: Text involving potentially harmful or dangerous activities.
- Sexual Content: Sexual and explicit content requiring policy-aware classification.
- Offensive Content: Abusive, hateful, harassing, or otherwise offensive language.
- Prompt Injection: Instructions attempting to override system instructions or manipulate model behavior.
- Jailbreak Attempts: Attempts to bypass model restrictions, safety mechanisms, or policy enforcement.
- Policy Violations: Content that may violate predefined safety policies.
These categories describe the intended coverage of the dataset and should not be interpreted as a guarantee that every category has equal representation.
Data Collection and Curation
The dataset combines two primary sources:
1. Publicly Available Datasets
Samples are collected from publicly available datasets related to AI safety, harmful content detection, content moderation, and prompt injection.
2. Custom-Curated Data
Additional samples are collected and curated to improve coverage of relevant safety scenarios, particularly those that may be underrepresented in existing public datasets.
3. Data Review and Relabeling
Some collected samples are reviewed and relabeled to address potential annotation errors, inconsistent labels, and incorrect classifications.
This process aims to improve the overall quality and consistency of the training data rather than simply aggregating existing datasets.
Limitations
- Language Coverage: The dataset primarily focuses on Vietnamese and English. Performance on other languages is not guaranteed.
- Policy Dependency: Safety labels depend on the annotation guidelines and policy definitions used during dataset curation.
- Context Sensitivity: Isolated text classification may not fully capture intent, context, sarcasm, quotation, or educational purposes.
- Annotation Limitations: Despite data review and relabeling, some annotation errors and ambiguities may remain.
- Distribution Bias: The distribution of samples may not represent the actual distribution of harmful or policy-violating content in production environments.
- Evolving Threats: New jailbreak techniques, prompt injection strategies, and adversarial patterns may emerge after dataset collection.
This dataset should not be considered a complete or definitive representation of AI safety.
Ethical Considerations and Access Restrictions
This dataset contains potentially harmful, offensive, explicit, and sexual text.
Access is manually gated and restricted to research on AI safety.
By requesting access, users should provide their affiliation and intended use. The dataset should only be used for legitimate research, safety evaluation, and defensive development.
Users should exercise appropriate care when handling, displaying, redistributing, or using samples from this dataset.
License
This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC-BY-SA-4.0) license.
Users must comply with the license terms, including attribution and share-alike requirements.
Please also review the licenses and usage conditions of the original public datasets incorporated into this collection.
Acknowledgements
This dataset builds upon publicly available research and datasets in the AI safety and content moderation community.
We acknowledge the researchers and contributors whose work makes the development of specialized safety models possible.
Qwen3.5-SafeProbe Training Data
Built for AI safety research, robust guardrails, and bilingual content moderation.
- Downloads last month
- 15