extract/embeddings/ec_classification/ref_composition/meta.json

{
 "encoder": "ref_composition",
 "family": "non-learned reference",
 "weights": "none",
 "n_parameters": 0,
 "rule": "20 amino-acid frequencies, log length, length/600, fraction outside the 20-letter alphabet",
 "inputs_passed": "amino-acid sequence only",
 "hidden_size": 23,
 "task": "ec_classification",
 "ids_copied_from": "esm2_150m",
 "n_per_split": {
  "train": 6000,
  "valid": 1500,
  "test": 3000
 },
 "purpose": "downstream floor: what the task scores without any learned representation",
 "date": "2026-09-21 14:00"
}