Only what we created: adapter 39MB, head 230KB via LFS. Base Qwen model
stays a from_pretrained download - consumers fetch the canonical copy.
predict.py tested on the 4090 (soccer 0.96 / flu_report 0.977 / yearbook 0.94).
example-repro: spec-faithful Qwen3-Emb+LoRA+prototype-head training on synthetic data
Sasha Malahov
68a82f9
Reimplements the jev-topic-head training from the public spec checklist
(no original code available): 56-anchor prototype head seeded from topic
descriptions, LoRA r16, two AdamW groups, warmup->decay, 1/sqrt(count)
sampling with hard-pair x2, early stop on val macro-F1, 3 seeds + eval.
Synthetic structural twin of the private dataset; real-data rerun pending.
