A framework to measure how much of an output originates from user input (prompt), training data biases, inductive biases from model architecture, or novel composition of retrieved information.

1. Introduction

This project implements a Mirror-Aware Inference that performs "bias-tracking" by analyzing the model's internal state during generation. The scripts perform a series of backpropagation passes to measure the influence of different components… See the full description on the dataset page: https://huggingface.co/datasets/ronniross/mirror-aware-inference.