cxj05h/laya-ledger-mlx

laya-ledger (MLX, FP16)

A small fine-tune of Laya (Convai Innovations, Apache 2.0) for the routine decisions of a work ledger: is the user asking to close a row, move focus, or add a new row - and which existing row do they mean? It is the optional local decider of the project-ledger and castai-sessions Kimchi packages. Native MLX FP16, for Apple silicon, loaded with laya_mlx.load("<this repo>", dtype="float16").

Modified from the base model: fine-tuned for 4 epochs with the upstream typed-decision recipe (RLCD + per-type temperature calibration) on 2,400 synthetic ledger sessions (10 invented "departments", 160 invented work items; every person and company is fictional). No real user data. Converted to MLX with laya-mlx convert.

What it is for, and what it is not

• Its row answer is calibrated: when its top row scores 0.80 or more it was right 91-93% of the time on our evaluations (small samples: 11 and 14 cases); below that, ask the user.
• Yes/no questions (close / move focus / new item): about 90-95% on our evaluations.
• It is an advisor, not an actor. It is not a general language model and makes no claim outside this task.

License and credit

Apache License 2.0, the same as the base model. Base model: Laya by Convai Innovations (https://huggingface.co/convaiinnovations/laya); MLX conversion tooling: laya-mlx. This repository contains modified weights.