How I shrunk a Russian embedder to 24 million parameters — and almost ruined it with one digit
An engineer trained a lightweight Russian dense retriever, STRIZH, with 24.4M params and 4 layers, for local RAG on shared GPU. A single wrong number in tokenizer config (model_max_length=256) dropped Recall@10 from 0.589 to 0.448, nearly voiding the work. The model is meant for fast first-stage retrieval on AMD Strix Halo, competing with bge-m3 in co-residency scenarios.
DeepPavlov
BAAI
Microsoft
Habr — хаб ИИ28.07 · 20:02