Structured tool-calls from a 25M model

A 26-million-parameter ternary model — tiny — fine-tuned to turn a plain request into a structured tool call: {action, title, due, priority}. The interesting part is that it generalizes: on requests it never saw in training, it picks the right action, due, and priority about 98% of the time — evidence that a model this small genuinely learns the routing, rather than memorizing a fixed list.

Caveat. A small, under-trained research prototype. It’s strong at the structured/classification part (action · due · priority), but weak at copying an arbitrary title back verbatim when the wording is unfamiliar (~23% on novel phrasings) — open-vocabulary copy is the hard part for a model this size. Illustrative, not production.

A request

Or try one

Remind me to call the dentist on…Add a low priority task to prepa…New task: water the plants this …Create an important task to subm…Remind me to charge the camera b…

raw

Method

The base is a 26M-parameter ternary decoder pretrained on a curated text corpus, then fine-tuned (response-only) on request→tool-call pairs. At inference it greedily decodes the JSON for the request. There’s no grammar constraint or post-processing here — the model emits the structure directly; a real deployment would constrain decoding to the schema. What this shows is that the capability — map intent to a structured action and classify its fields — fits in a model small enough to run on a phone, and it generalizes to new requests.