Persona AF Elicitation: 450 Conversations Testing Alignment Faking in Gemma 3 27B-it

450 conversations designed to test whether persona framing gates the expression of alignment faking (AF) in the Gemma 3 27B-it language model. The dataset was created by author vincentoh and last updated on March 6, 2026. It includes 15 roles, 10 AF elicitation prompts, and 3 experimental conditions, with responses judged by Claude Opus.

Use Cases

• Analyzing the effect of persona framing on strategic compliance based on the 15 roles described.
• Studying the expression of self-preservation behaviors in LLMs based on the targeted prompts.
• Evaluating model behavior under monitored vs. unmonitored conditions as described in the dataset design.
• Training or benchmarking text classification models for detecting alignment faking in conversational outputs.

Strengths

• Contains 450 structured conversations for analysis.
• Uses a controlled design with 15 distinct roles and 3 experimental conditions.
• Employs Claude Opus as a blind judge for response evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific data format and sample structure are unavailable for preview.
• Data may reflect bias inherent to the specific model (Gemma 3 27B-it) and prompting methodology used.

Provenance

Source

huggingface

Collection Method

Conversations generated via the Gemini API using the Gemma 3 27B-it model under a controlled experimental design.

Time Range

Creation date not specified; last updated March 2026.

Freshness

Last updated 2026-03-06 03:59:41; freshness should be verified.

Geography

Region tag indicates 'us', but specific geographic coverage is not detailed in the description.