Pre-generated feature labels and calibration data for Neuroscope, an SAE-instrumented LLM inference server.
Auto-interp labels — human-readable descriptions for SAE features, generated by running max-activating examples through an LLM
Calibration stats — per-feature firing rates and activation statistics, used for… See the full description on the dataset page:
https://huggingface.co/datasets/cjroth/neuroscope.
