bitlabsdb/bad-classifier-mistral-7b-fairsteer

FairSteer BAD Classifier (Secure)

Biased Activation Detection (BAD) classifier optimized for mistralai/Mistral-7B-Instruct-v0.3. This model detects whether the LLM's internal activation (at layer 25) indicates biased reasoning.

This repository contains only SafeTensors weights for security.

Model Details

• Base Model: mistralai/Mistral-7B-Instruct-v0.3
• Target Layer: 25
• Architecture: Linear Probe (Dropout -> Linear)
• Performance: 75.19% Balanced Accuracy

Artifacts

• model.safetensors: Weights (SafeTensors only)
• scaler.pkl: StandardScaler (Required for inference preprocessing)
• config.json: Architecture configuration

Usage (FairSteer)

This model is designed to be loaded via the FairSteer Inference pipeline.