maximuspowers/muat-fourier-3

Subject Models for Interpretability Training These examples are intended for training an interpreter to: Identify what patterns a model classifies as positive based on an activation signature, with examples of: trained model + signature → pattern identification. Signature Extraction Neuron Profile Methods fourier Prompt Format separate Signature Dataset dataset_generation/exp_1/signature_dataset.json Model Architecture Number of Layers 4 to 6… See the full description on the dataset page: https://huggingface.co/datasets/maximuspowers/muat-fourier-3.