Aegis-Safety-DPO: PolarAI's Safety Alignment Preference Data

Aegis-Safety-DPO is a manually-curated preference dataset designed for Direct Preference Optimization and Group Relative Policy Optimization. Created by PolarAI, the dataset focuses on training models to refuse malicious requests rather than provide preachy or evasive responses. The dataset was last updated on March 1, 2026.

Use Cases

• Training models to refuse malicious requests based on the described safety alignment approach.
• Direct Preference Optimization (DPO) based on the high-density, manually-curated preference data.
• Group Relative Policy Optimization (GRPO) based on the described dataset design.
• Analyzing model refusal behavior based on the Chain-of-Thought (CoT) analysis mentioned in the description.

Strengths

• Dataset is described as 'high-density' and 'manually-curated'.
• Dataset is designed for specific training objectives (DPO, GRPO) rather than generic safety responses.
• Dataset was last updated on March 1, 2026.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

PolarAI

Collection Method

Described as (mostly) manually-curated.

Freshness

Last updated 2026-03-01 17:11:37.