bev-decider-0.4B: an open-source decision model with a Jev-compatible API
A 0.4B causal decision model built on the first 20 layers of Qwen runs in about 50ms per parallel forward pass and ships with a Jev-compatible /v1/systemone API, with code, weights and data all open source.
Releasing bev-decider-0.4B: a lightweight causal decision model that is architecturally invariant of input choice order. Its built on the first 20 layers of Qwen. Processes the prompt and choices in one parallel forward pass. Runs superfast at ~50ms even on my 3-yr old mac (no speedups in the demo). This started as for a ed youtube video, but I was having too much fun, so I'm releasing everything. Dataset, code, weights are all open source. The video (full project, start to finish) drops this week! Bev is pip installable and comes with a Jev-compatible /v1/systemone api. You can find benchmark details in the HF model page. :) Github: https://github.com/avbiswas/bev-train Model: https://huggingface.co/avbiswas/bev-decider-0.4B Data: https://huggingface.co/datasets/avbiswas/bev-decision
This post quotes another post: x.com/neural_avb/status/2103822851359027207