animalialive/nalu-decisions

Nalu Decisions

A small on-device decision model for the Nalu app. It reads one message and picks what the assistant mainly needs to handle it: none, remember, web, calendar, reminder or draft. The question it was trained on is in question.json; send it exactly as written. When the person's previous message is known, the input follows state_format.json (Previous message: … / Message: …); otherwise it is the message alone.

It also answers four more questions, written in questions_v3.json (send them exactly as written): service (which connected service a message is about; the options are the services' own descriptions), depth (quick, task or heavy), tool (which of up to 12 tool cards a request needs first) and relevant (yes/no: does a fact about the person help answer the message).

Fine-tuned from the multilingual Laya checkpoint (convaiinnovations/laya) with laya-train, then converted to MLX float16.

Scores

set	accuracy
held-out 90 (own messages)	0.978
MASSIVE test slice	0.949
original 36	35/36

question	set	accuracy
service	services never seen in training	0.900
depth	own messages	0.927
tool	own messages (top choice)	0.880
relevant	own message and fact pairs	0.790

Credits and licenses

• Laya, by Nandha Kishor M and ConvAI Innovations, Apache-2.0. The upstream LICENSE is included.
• Training data includes MASSIVE 1.1 from Amazon, CC BY 4.0 (FitzGerald et al., 2022), plus synthetic messages written for this project.
• This model is released under Apache-2.0.