README.md

metadata

pretty_name: Canonical embedding sequences
task_categories:
  - feature-extraction
tags:
  - protein
  - sequences
  - embeddings
  - esmc
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: sequences.parquet
dataset_info:
  features:
    - name: seq_sha256
      dtype: string
    - name: sequence
      dtype: string
    - name: full
      dtype: bool
  splits:
    - name: train
      num_examples: 3551562

Canonical embedding sequences

Summary

The normalized protein sequences required by the canonical embedding stores. One row per SHA-256 sequence key, with a single boolean indicating whether full per-token embeddings are required. Every row requires pooled embeddings. This sequence list has no supervised labels, pairs, source-membership columns or biological safety classification.

Quantity	Count
Sequences	3,551,562
Full per-token embeddings required	852,475
Pooled only	2,699,087

Homology leakage

No homology split is defined for this unlabeled inventory. The viewer's train split contains the whole inventory and supplies no model evaluation or leakage guarantee.

How it was built

1. Retained every protein in the previous sequence inventory, preserving its normalized text and key.
2. Read all protein columns of the selected datasets at immutable revisions, including both chains of pairs.
3. Removed ASCII whitespace and uppercased letters without rewriting residues; deduplicated by SHA-256.
4. Combined full-embedding requirements with boolean OR and verified all source coverage and written rows.
5. Appended the proteins of the Atlas v3 interaction corpus the list lacked, and required full per-token embeddings for every one of its proteins.

Columns

Column	Meaning
seq_sha256	SHA-256 of the normalized ASCII sequence; stable embedding-store key
sequence	Full uncropped normalized protein sequence
full	True requires full per-token embeddings as well as pooled embeddings; false requires pooled only

Example rows

Use the viewer for actual rows. Sequence length is derived from sequence; per-source flags are omitted.

What a model trained on it may say

This inventory has no target and supports no binding, function, affinity or safety claim. Embedding qualification, truncation and model-specific numerics belong to the embedding profile.

Sources and licence

Sequence-only redistribution follows the existing public sequence-inventory authorization. Source terms remain applicable; this list grants no new licence for source labels or measurements. Source revisions and membership audits are retained separately from this public sequence-only release.

Files

sequences.parquet is the single data file. The repository rename preserves historical Git revisions.