SAE Release for Zeitgeist (under review): Chronos-2, encoder block 2, T-SAE (4x, k=4)
Place both folders in the root of the accompanying code; all stored paths are relative to it.
checkpoints/.../final/cfg.json: training configcheckpoints/.../final/inference/: SAE weights and config, core metrics,feature_analysis.h5(all analysis results)data/synthetic_timeseries/: known-concept dataset
Report: run features analyze --sae-path checkpoints/chronos2-baseline-ks1m-gift5k/tsae_layer2_4x_k4/final/inference. This needs no GPU and no corpus.
Atlas: first regenerate the activation corpus with prepare-activations (about 140 GB). Its data sources are pinned in the code, so it rebuilds row for row and is checked against a stored fingerprint.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support