bev-train

What it is

Train a network that answers choice, yes/no (noul), and score questions using a Qwen backbone and a small attention head. Choices share the same starting position IDs, and the attention mask prevents each choice from seeing the others. The head scores the choices without depending on their order.

Getting started

From the README:

uv sync

Where the decision sits

bev-train runs its own weights or its own runtime rather than calling a hosted API, so its decisions do not depend on a vendor being up or on a key being funded. The trade is that the model is now yours to keep current.

How active it is

Created 2026-09-27, last pushed 2026-10-04, 83 stars and 14 forks when we read it on 2026-11-02. Everything in this category is weeks old at most, so a stale push date means something different here than it would on a mature project. Check the repository before you depend on it.