Synthyra/esmc_embeddings_staging

Synthyra/esmc_embeddings

Embeddings of the public ESMC models, computed under an embedding profile: a pinned container, pinned numerics, one fixed batch shape per protein and a pinned SAE readout, so every row is a function of its protein alone and bitwise reproducible on the hardware class that made it (Hopper: GH200, H100, H200). Each store is a FastPLMs feature store at <model>/<kind>/, the one canonical place for that model and kind; the profile column names the profile it was computed under, and a store computed under a new profile replaces it there. catalog.json lists them all. Each store holds every protein of one revision of its sequence list: a pooled store every row, a token store the rows flagged full, which need per-token rows. The sequences column names that revision (and the flag), and the store's coverage.json records the exact check that passed it: every protein's rows present, and a sample of proteins recomputed in new batches, equal bit for bit.

store	model	kind	proteins	GiB	streams	profile	sequences
esmc_600/pooled	ESMC-600	pooled	3,480,520	121.2	final_mean_var, layer_mean_var, sae_max, structural_mean_var	embedding_profile_v2	77db5440
esmc_600/token	ESMC-600	token	775,186	1,455.1	final_hidden, final_mean_var, layer_hidden, layer_mean_var, sae_codes, sae_max, structural, structural_mean_var	embedding_profile_v2	77db5440 full

stream	per protein
layer_hidden	(l + 2, d) bfloat16: the hidden state at the SAE's layer
final_hidden	(l + 2, d) bfloat16: the last layer, after the final norm
structural	(l + 2, 256): the structure fusion, a learned mixture of every layer projected to 256
sae_codes	(l + 2, 64) indices and values: each token's top 64 SAE codes, stored sparse
sae_max	sparse (c,) float32, c the SAE codebook: each code's maximum over the tokens
final_mean_var, layer_mean_var, structural_mean_var	(2w,) float32: the mean then the variance over the l + 2 tokens

A row is keyed by the SHA-256 of the uppercased, whitespace-stripped sequence. Per-token streams hold l + 2 rows per protein (row 0 CLS, rows 1..l residues, row l + 1 EOS, cropped at 2046 residues); pooled streams hold one vector.

Each store carries index/rows.parquet (per protein: the segment, part and row of each stream, sorted by key) and index/parts.parquet (per file: path, rows, bytes, SHA-256). A few proteins are read by HTTP range requests, so only their bytes are downloaded however large the store is; a whole store is a folder download:

from foundry.embedding.hub import read_rows

rows = read_rows("Synthyra/esmc_embeddings", "esmc_600/pooled", ["MKTAYIAKQR"], revision="<commit>",
                 streams=["sae_max", "final_mean_var"])
vector = rows["MKTAYIAKQR"]["final_mean_var"]  # a (2d,) float32 tensor

from huggingface_hub import snapshot_download
snapshot_download("Synthyra/esmc_embeddings", repo_type="dataset", allow_patterns=["esmc_600/pooled/**"])

Without the workspace: filter index/rows.parquet on sha256; each part is a safetensors file, whose first 8 bytes give its JSON header's length and whose header gives each tensor's byte span. A dense stream's row r is row r of values; a per-token or sparse stream's row r spans offsets[r]:offsets[r + 1] (indptr for sae_max) of values (and indices).