Evo 2 sparse autoencoder activations for the Escherichia coli K-12 MG1655 genome

Whole-genome sparse-autoencoder activations for every base of the Escherichia coli K-12 MG1655 chromosome (RefSeq GCF_000005845.2, NC_000913.3), computed from the blocks.26 embedding of Evo 2 7B with the Goodfire BatchTopK SAE, stored as a CSR matrix of 4,641,652 rows x 32,768 features. Includes the embeddings, annotation and protein structures used by an accompanying hands-on notebook that reproduces the E. coli panels of Figure 4 of the Evo 2 paper. See README.md in the files below for the full file list, the byte layout of the CSR files, and how to read one region over HTTP Range without downloading 13.6 GB. Version 3 rebuilds condition (ii) of Extended Data Fig. 7c in scramble.pt: the constant scrambled direct repeat is now drawn from the consensus of the thirteen repeats rather than from the leftmost copy, which carries three substitutions. The other fifteen embeddings in that file are byte-identical, and the other files are unchanged.