This is the shortest path from an empty machine to a typed decision you can act on. It assumes you have a GPU available and Python installed.
Step 1 — Install and pin
pip install laya
pip freeze | grep laya # 记录确切版本Pin that version. The interface moved through several minor releases in the first days after launch, and reproducing your setup later is much easier with a locked version.
Step 2 — Choose your checkpoints
Three checkpoints exist: an English one (~421M), a multilingual one (~322M) and a router. Decide based on your traffic:
- English only — load the English checkpoint
- Mixed languages — load the router plus both encoders
- Edge device — load one encoder only
Loading everything costs roughly 2 GB of memory. Use preloading to avoid cold-start latency on the first request.
Do not skip the router if your input is multilingual. The English checkpoint does not degrade gracefully — it returns confident, incorrect answers on non-Latin scripts.
Step 3 — Model your first decision
Do not start by writing code. Start by writing the question down in one sentence, then decide which primitive it is:
- Is the answer one of a short list? → choice
- Is the answer a number? → score
- Is the answer yes or no? → noul
If your sentence contains "and", you probably have two questions. Split them: each gets its own calibrated threshold, and you can tune one without retraining the other.
Step 4 — Get the call signature from the README
The concrete call signature has changed since launch, so copy it from the repository README rather than from any blog post. What stays stable is the shape of the thinking:
- You pass the text to be judged
- You pass a typed question describing the decision
- You receive values plus probabilities, not prose
There is no text to parse, which means no JSON repair loop and no regex against a model's rambling.
Step 5 — Check the confidence before you trust it
Run a handful of examples through and look at the confidence values. If everything comes back above 0.9, that is a warning sign, not a good result — the checkpoints ship overconfident.
Plan to temperature-fit on your own labelled data before any automated thresholding. That is the difference between a probability you can gate on and a number that merely looks like one.
Step 6 — Measure on your own hardware
The reported ~33 ms median is a T4 measurement. Measure p95 on the hardware you will actually deploy to, including:
- Cold start with preloading disabled
- Warm single request
- Batch throughput at your real batch size
Record these before you write any capacity plan.
What a first run should tell you
A useful first session ends with three things:
1. One working typed decision on your own text 2. A p95 latency number measured on your target hardware 3. An honest note about whether the raw confidence values look usable — they usually do not, and that is expected
From there the work is fine-tuning and calibration, which is where the real accuracy comes from.