A ~0.4 local model to turn typed questions into structured decisions - BaseDecision
I’ve been working on BaseDecision - Turn typed questions into truth-based decisions.
Give it some text, a question, and possible answers. It picks an answer. Intent classification, routing, yes/no checks, ratings - that’s the job.
I built this model to go toe-to-toe with the frontier models.
My goal was simple & ambitious: make a small, local model competitive with much larger models on these focused tasks. For me, System‑1 decisions need both speed and accuracy. Otherwise, why bother building a small model?
If we didn’t want speed we could just use ChatGPT.
It outperforms pretty much every model of its size in the field.
After roughly 320M additional training tokens, BaseDecision led 5 of 8 benchmarks in a third-party comparison against Laya, two GLiNER variants, and Decision 1.0 Kai 0.6B. Full results are in the repo, including where it loses.
Fellow open source devs, try the model, developed a package for easier access to the model.
Roast it if it deserves it.
2-2.5GB free ram is all you need.
I will take your feedback seriously and deliver something amazing that’d take on OpenAI, TypeSafeAI soon enough.
https://huggingface.co/onlyaady/BaseDecision
https://github.com/hrudayaditya/BaseDecision