Document embedding provenance and source terms

NOTICE

@@ -8,6 +8,8 @@ datasets retain their original licences and attribution requirements:
 * GATA1: Martin-Rufino et al., GEO accession GSE215253; processed object from
   the PerturbNet repository (cyclopeta/PerturbNet_reproduce).
 * JAK1: Cooper et al., ENA PRJEB48915; Zenodo record 10418435.
+* Molecular features: FAIR ESM-1v model (Meier et al., 2021); only derived
+  embeddings are included and model weights are not redistributed.
 
 The repository-level Hugging Face metadata therefore uses `license: other`.
 No repository-level Apache-2.0 grant is made here, and the derived release

README.md

@@ -47,7 +47,9 @@ benchmark. The underlying scSNV-seq data are available from
 ### Molecular features and compact response data
 
 - `allele/esm1v_embeddings.npz`: ESM-1v variant embeddings used as molecular
-  features; no model weights are included.
+  features; no model weights are included. The embeddings were generated with
+  the FAIR ESM-1v model (Meier et al., 2021; see the
+  [FAIR ESM repository](https://github.com/facebookresearch/esm)).
 - `allele/real_deltas.npz`: compact variant-level response profiles derived
   from the processed allele datasets. This is not a cell-level matrix.
 - `allele/barcode_rebuild_provenance.json`: provenance, source checksums and