- Laya is a system that receives texts in over 100 languages and performs multi-choice format decision-making through a single neural network process.
- The main models consist of an English model, a multilingual model for 100 languages based on mmBERT, and task-specific typed-decision models, with a router that automatically selects the appropriate model based on the request.
- It can process document inputs up to 8192 tokens, and for long documents of around 4000 tokens, it correctly answers 16 to 18 out of 20 test requests (measured based on publicly available research scripts).
- It can be installed in Python environments 3.10 or higher, supporting reinforcement learning, ONNX conversion, and batch processing. Separate libraries for TypeScript and Node.js are also provided.
For sorting incoming questions into a fixed set of categories, this looks worth trying instead of asking a model to generate a long answer. Being able to handle several languages in the same workflow is appealing too. I'd still want to see what happens to speed and accuracy with longer documents. A few real Korean documents would tell me more than the short examples.