kenga-embed-z

Older non-instruct Z-Embed, seq 96. Best handmade-holdout file we still have for Z (12/12 EN, 11/12 RU).

Do not apply the Giga instruct prefix to this checkpoint. Encode raw UTF-8 bytes.

from modeling_kenga_embed import KengaEmbed
m = KengaEmbed.from_pretrained("GermannM/kenga-embed-z")
# instruct=false in config
q = m.encode_queries(["Where is the capital of Russia?"])  # raw
d = m.encode_documents(["Moscow is the capital of Russia."])

Holdout (not ruMTEB): EN 1.000, RU 0.917, miss = Russian H2O.

RuSTS on this file (same weights, 2026-09-02, no instruct): 0.3723.

Later Z last weights are instruct+seq160 and currently lose the holdout if you mix templates. This upload is the frozen best.