Train Laya With Claude Code: The Free Model That Beats Jev

Laya is a free, open source (Apache 2.0) project by NandhaKishorM. You give Laya a piece of text, like a support ticket, plus a few choices, like Billing, Technical or Other. Laya picks one in a single pass, with no text writing, so it answers in about 30 milliseconds on a GPU. It is built to do the same job as Jev, the hosted model from TypeSafe. Laya is an independent project, not an official open source copy of Jev.

Speed: Laya's own benchmarks list 32.8 milliseconds for one question on a T4 GPU. Jev's published speed is 236 to 276 milliseconds. Those Jev numbers come from other people's measurements, so treat them as a rough guide. Speed also depends on your computer: a newer GPU is faster, and a laptop CPU is slower.

Accuracy: Straight out of the box, Laya scores 36 percent on the typed-decisions test, against 73 percent for Jev. Laya starts weak because the base model has not seen your kind of questions yet. After training on that test's examples, the Laya checkpoint called laya-typed-decisions scores 77 percent. That is the author's own result on one benchmark, so test it on your own data before you rely on it.

How to install Laya: copy the repo URL below, paste it into Claude Code, and say: install this.

Or install Laya by hand. It needs Python 3.10 or newer:

The Laya repo ships a fine-tuning notebook for Kaggle (two T4 GPUs) and a script for Apple Silicon Macs. Both run the whole loop: build a dataset, train, calibrate and test. Your job is to collect examples where you already know the right answer, like 500 old support tickets with the department that handled each one. Claude Code does the rest.

Best practices for training Laya:

• Use real examples from your own work (tickets, emails, reviews) with the correct answer already attached. Laya learns from your answers, so wrong labels teach it wrong habits.
• Keep each question to about 20 choices or fewer. Laya's own docs say Jev is still stronger on very long option lists.
• Stay in English with the English checkpoint. For other languages, use laya-multilingual, because the English checkpoint gets much worse outside English.
• Keep some examples out of training and test on them. A score on examples Laya already saw tells you nothing.
• Expect over-confident answers. Laya's calibration step fixes this, so run it before you trust a confidence number.

Prompts to try with Laya:

• Pick one job where you already have answers, like sorting support tickets by department.
• Run untrained Laya on 50 of your examples first. That gives you a before score.
• Train on the rest, then score Laya again on the 50 you held back. The gap between the two scores is your real result.