Gretel Safety Alignment Dataset with 8,361 Prompt-Response Triplets

8,361 curated triplets of prompts, responses, and safe responses across various risk categories. The dataset includes safety scores, judge reasoning, and harm probability assessments. It was created by Gretel.ai and is available under the Apache License 2.0.

Use Cases

• Fine-tuning language models for safety based on the provided safe responses.
• Benchmarking model safety performance using the included safety scores and harm probability assessments.
• Analyzing failure modes in model outputs based on the judge reasoning annotations.
• Generating synthetic training data for safety classifiers using the prompt-response triplets.

Strengths

• Contains 8,361 curated entries, providing a substantial base for analysis.
• Each entry is a triplet with a prompt, a response, and a safe response, enabling comparative safety training.
• Includes multiple safety-related annotations per entry: safety scores, judge reasoning, and harm probability assessments.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the total data size and sample structure are unknown.
• Last update date is unknown; freshness unverified.

Provenance

Source

Gretel.ai

Collection Method

Curated and synthesized, likely for safety alignment research.

Time Range

Temporal coverage is unknown.

Freshness

Last updated date is unknown.

Geography

Spatial coverage is unknown.