JevEmbed publishes a Qwen3-Embedding-4B fine-tune for Choice/Score/Noul
On 28 September 2026, HIT-TMG released JevEmbed-Qwen3-Embedding-4B: a LoRA fine-tune of Qwen3-Embedding-4B merged into a Sentence Transformers checkpoint for Choice, Score and Noul decisions. It joins earlier 0.6B and KaLM weights in the JevEmbed stack (Python API, CLI, optional Jev-shaped HTTP server).
Training used one epoch on the 1.6M-question JevEmbed-Data split. Author-reported (UnverifiedClaim): 85.86% overall hard-label accuracy on the project’s own test split versus 36.29% for the untuned base embedder under the same prompts — an in-domain figure, not a zero-shot typed-decisions or JevBench score.
The framework can also wrap frozen embedders; this news is about the trained 4B checkpoint. Catalog entry: /models/jevembed. Nearby: Laya, GLiNER2.5-Decide, Julia-1.