codepawl/tacet-sonata
Text Classification • 0.1B • Updated • 18 • 1
AI & ML interests
Small, fast encoder models for typed decisions with calibrated probabilities. Multilingual NLP, open weights.
Recent Activity
Organization Card
CodePawl
CodePawl is a small software studio in Vietnam. We build small encoder models that answer typed questions and give a probability for every option, and we release them as open weights under Apache 2.0.
Tacet Sonata
Our first model is Tacet Sonata. You send it a piece of state, like a support thread, a security alert, a vendor invoice or an agent trace, plus a few typed questions. A question can pick one of N options, place the state on a scale, or answer yes or no. It reads the state and every question in one encoder pass and returns a probability for each option. It never generates text, so there is nothing to parse.
On the LocalLLaMA typed decisions test split it matches Laya on accuracy at a third of the size, and on one NVIDIA L4 it serves 14x the requests per second. It is strongest on inputs shaped like the benchmark's four workflows. Free text, dates and counting are still weak spots.
- Code,
pip install tacetand a local server: github.com/codepawl/tacet - Site and docs: tacet.codepawl.com
- Questions and ideas: Discord
datasets 0
None public yet