What is Laya

Laya is a small open model from ConvAI Innovations for decisions you would rather not ask a chatbot to write out. You give it a document — an email, a ticket, a JSON blob — and a list of questions whose answers are already typed. It returns those answers with probabilities in one forward pass.

The authors describe three question shapes:

• choice picks one option from a list you supply at request time.
• score places the text on an ordered rubric.
• noul answers a yes-or-no question with P(true).

Because the output is a distribution over options you defined, there is no paragraph to parse. That is the product. It is not a general chat model.

The three checkpoints

The hub repo holds the family. ConvAI’s card describes them this way: an English 421M checkpoint, a multilingual 322M checkpoint for a wide set of languages, and a 421M typed-decisions checkpoint specialized on four workflows. A router in the Python package can send a request to the English or multilingual weights after a quick script check.

What the authors say it is not

The same card is blunt about limits. The base checkpoints are weak zero-shot on the typed-decisions suite; the 0.766 figure belongs to the fine-tuned checkpoint. Choice questions with dozens of labels get squeezed by a shared token budget. Probabilities need a temperature fit on your own data before you trust them. Use the multilingual checkpoint when the text is not English.

Their longer essay is at laya.convaiinnovations.com. Weights are Apache 2.0. This site does not ship weights and does not speak for ConvAI.