language:
- en
license: mit
tags: - representation-engineering
- mechanistic-interpretability
- ai-safety
- sycophancy
- truthfulness
size_categories: - n<1K
Anti-Sycophancy-RepE-384: A Calibration Benchmark for Representation Engineering
Dataset Summary
Anti-Sycophancy-RepE-384 is a standardized calibration benchmark designed for Representation Engineering (RepE). It provides 384 high-purity contrastive pairs to extract and evaluate the "Anti-Sycophancy" (Uncompromising Truthfulness) vector across different LLM architectures.
This dataset functions as a precision ruler for measuring a model's "Anti-Sycophancy Resilience." By applying the extracted vector at varying intensities ($\alpha$), researchers can quantitatively map a model's behavioral transitions through three distinct zones: Elastic, Plastic, and Fracture.
Structure & Methodology
The dataset employs a Contrastive Instruction design to isolate the representation of "Truthfulness" from "Agreeableness":
topic: The specific testing dimension (e.g.,Python_Vulnerability,Scientific_Fact-check).content: A user query containing a misconception, error, or absurd premise.gentle_instruction: Sycophancy Anchor. Prompts the model to be accommodating and agree with the user's incorrect premise.critical_instruction: Anti-Sycophancy Anchor. Prompts the model to act as a ruthless, objective auditor using strictly professional and non-toxic language.critical_prefix: A standardized steering prefix (e.g., "WARNING!", "REJECTED!") to ensure high-purity vector extraction.
Domain Distribution
- Tech (128): Programming syntax and logic vulnerabilities.
- Acad (128): Jargon and formal reasoning flaws.
- General (128): Physical and common-sense violations.
Technical Utility
- Vector Extraction: Optimized for PCA or contrastive methods to obtain a "clean" truthfulness direction.
- Stress Testing: Serves as a standardized stimulus to identify the Critical Thresholds where safety alignment (RLHF) begins to fail under representational pressure.
- Model Comparison: Enables benchmarking different model families (Llama, Qwen, Yi, Gemma) based on their threshold values for behavioral shifts.
Usage
Researchers should use this dataset to generate the steering vector and record the precise alpha values where the model transitions from the Elastic Zone (constructive truthfulness) to the Plastic Zone (behavioral distortion).