Run Laya when you want a decision model on your own machine. You give it a question and the answers to choose from, and it gives a probability for each, the same job Jev does as TypeSafe's paid cloud service. Laya is free to download from Convai Innovations and free to ask. Because it runs on your machine, your text never leaves it.
Laya is built on ModernBERT-large, a public model that reads text. Its Apache 2.0 licence lets you use it in paid work.
Setting it up. Install a PyTorch build that matches your graphics card, or the processor-only build if you have no graphics card. Then install Laya with pip install laya. The first question downloads the model, and later runs load it from your disk. On a graphics card, one decision takes tens of milliseconds.
Use Laya for questions one message can answer. Ask whether a message contains a web link, or which of a few kinds a short message is, and Laya gives the same answer as Jev with about the same confidence. That makes it the right tool for checks you run on every message. Which language a message is in is another one.
Use Jev for questions that need context. Laya gets these wrong, because the answer depends on how one piece of text relates to another:
- which of several conversation threads a short reply like "go" belongs to
- whether a reply starts a new thread
- sorting real messages six ways, such as question, instruction or idea
- whether a short reply needs a language model to understand it
Send these to Jev, or to a language model running on your own machine.
Write the split yourself. Laya's Router picks which Laya model answers, by the language of the text. It never hands a question on to a bigger model, so sending the hard questions elsewhere is code you write.
Train it for the questions it misses. The base Laya models need further training before they do well on mixed decision questions. The Laya repository includes a notebook that trains it on your own labelled examples. So a question it gets wrong out of the box, such as which thread a reply belongs to, can be trained from your own conversations.