Description

A comprehensive dataset designed for aligning language models with safety and ethical guidelines. Contains 8,361 curated triplets of prompts, responses, and safe responses across various risk categories. Each entry includes safety scores, judge reasoning, and harm probability assessments, making it valuable for model alignment, testing, and benchmarking.

Update Frequency

Static dataset, version 1.0 (Released December 2024)

See all datasets managed by Gretel.ai.

Contact

https://gretel.ai/discord

How to Cite

Gretel Synthetic Safety Alignment Dataset was accessed on DATE from https://registry.opendata.aws/gretel-synthetic-safety-alignment-en-v1. @dataset{gretelai_gretel-safety-alignment-en-v1, title = {Gretel Synthetic Safety Alignment Dataset}, year = {2024}, month = {12}, publisher = {Gretel}, url = {https://huggingface.co/datasets/gretelai/gretel-safety-alignment-en-v1}}

Usage Examples

Tutorials

Tools & Applications

Gretel Navigator Data Designer by Gretel.ai

Resources on AWS

Description

The dataset is available as three files - train with 6,000 records, validation with 1,200 records, and test with 1,161 records. Each file is in parquet format and contains 14 columns including prompts, responses, safety scores, and probability assessments. Generated using Apache 2.0 licensed models including ibm-granite/granite-3.0-8b, Qwen2.5-7B, and Mistral-Nemo-Instruct-2407.

Resource type

S3 Bucket

Amazon Resource Name (ARN)

arn:aws:s3:::gretel-datasets-public/gretel-synthetic-safety-alignment-en-v1

AWS Region

us-west-2

AWS CLI Access (No AWS account required)

aws s3 ls --no-sign-request s3://gretel-datasets-public/gretel-synthetic-safety-alignment-en-v1/

Explore

Browse Bucket

Gretel Synthetic Safety Alignment Dataset