flax-embedding-2

Georgian sentence-embedding student distilled from a teacher embedding model by geo_distill, pushed by geo-distill train --push-to.

• student_config.json — architecture + provenance (which MLM encoder and tokenizer it came from)
• student_params.msgpack — the parameters of the best epoch, by held-out Spearman
• teacher_mean.npy — the training-split teacher mean; subtract it from a raw teacher vector to map it into this student's space
• student_state.msgpack / train_state.json — the newest checkpoint's latest parameters, Adam moments and position, so training continues across capped sessions

Machine-managed: every push replaces the previous snapshot.

Use it:

python -m geo_distill fetch-student ZurabDz/flax-embedding-2
python -m geo_distill eval --model-dir artifacts

Continue training it:

python -m geo_distill train ... --push-to ZurabDz/flax-embedding-2 --resume