Spaces:
Running on Zero
Running on Zero
A newer version of the Gradio SDK is available: 6.25.0
Offline study artifacts
This directory contains the small, publishable outputs from the completed FeatureLens study. Large activation matrices, model weights, SAE checkpoints, and task checkpoint markers are intentionally excluded.
Primary study artifacts:
feature_catalog.csvβ train-selected SAE candidates and held-out AUROC/F1.layer_metrics.csvβ dense-probe and SAE reconstruction diagnostics.stability.csvβ paraphrase stability measurements.selection_stability.csvβ 128-resample candidate-selection sensitivity.causal_results_final_token.csvβ final-token causal baseline.causal_results_max_active.csvβ max-feature-activation causal study.causal_position_summary.csvβ coverage and task-level position-sensitivity synthesis.feature_set_results.csvβ final-token top-1/3/5 feature-set diagnostic.study_feature_summary.csv/study_summary.jsonβ cross-concept evidence synthesis.summary.json/report.mdβ measured executive summary and report.split.jsonβ fixed train/held-out paraphrase-group split.figures/β report figures used by the public Study tab.
Run python -m scripts.validate_artifacts to validate the committed study bundle.