Data for: Ground-truth evaluation of sparse autoencoder features on spatial, single-cell and language-model embeddings

Derived data and results for the paper. All authors contributed equally. inputs__section_A.npz, inputs__section_B.npz: cell IDs, cell types and Expr50 / CellCharter / CadaST embeddings of the 50,000 CosMx cells analyzed per section inputs__pelka_geneformer.npz: Geneformer V2 embeddings and cell types of 12,881 Pelka et al. cells inputs__LLM_pythia70m_L3.npz: Pythia-70m layer-3 activations and token classes for 50,000 WikiText-103 tokens inputs__markers.json, inputs__counts_sha256.json: reference lineage markers; checksums of the count matrices used results__*: all run results and the final tables Count data are available from their original sources (see the code repository): https://github.com/AH-Shah/sae-groundtruth Version 1.1.0 adds results__results_v1.1.0_symmetric_consensus.zip: the runs and tables of the symmetric consensus used in the paper (*tsae_cl*, SYNTH_ROT_CL*, TABLES_v3cl.md, REAL_v3cl.csv, SYNTH_v3cl.csv, SYM_TEST*.csv); unzip it into data/results. Earlier files are unchanged. Version 1.2.0 adds results__results_v1.2.0_fine_types.zip: fine cell-type and within-lineage coverage of consensus and single SAEs and rotations on the five spatial embeddings and Geneformer, three seeds (FINE_TYPES_*.csv, FINE_TYPES_SUMMARY.md = Supplementary Table S6, and the scripts that produced them); unzip it into data/results. Earlier files are unchanged.