Laya-Finetune
Toolkit that turns Laya into a task specialist: LLM-generated and verified training data, RLCD training, calibration and longer context.
Overview
Laya Finetune turns Laya, the open System One decision model, into a specialist for your own task. You describe the decisions in a YAML file. One LLM writes training cases labelled by construction, a second LLM filters out wrong labels, and the tool trains with RLCD, calibrates confidence and can grow the context from 1,024 up to 32,000 tokens. Validation uses a hand-written test set that is never trained on, and the result is a checkpoint that loads with laya.load() and answers in milliseconds on your own GPU.
The README compares Jev 1.13 with Laya's multilingual and English models: Jev has closed weights, about 32,000 tokens of state context and 70 to 500 ms latency at $0.042 per million input tokens, while Laya multilingual has open 322M weights, 1,024 tokens of context and 10 to 30 ms on a desktop GPU, but out of the box misses more often, has uncalibrated confidence and struggles beyond 20 options. Fine-tuning on a single task is presented as closing most of that gap in four commands.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this ticket? (example task) | — |
| Score | How urgent is it? | — |
| Yes / no | Is the person blocked from working? | — |
How it works
- Describe the task and questions in task.yaml
- Generate labelled cases with an LLM
- Verify labels with a second LLM
- Train with RLCD, calibrate and extend context
- Validate on a hand-written test set and load the checkpoint
Strengths
- Specialises an open model for one task
- Calibration step included
- Context extension up to 32k
- Held-out hand-written validation
Caveats
- Needs a GPU and LLM API access for data generation
- Synthetic data quality bounds results
- Specialist checkpoints do not generalise to other tasks
Getting started
See the Quick start in the README: write task.yaml, then run the generate, verify, train and val steps.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- training pipeline (PyTorch)
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Fine-tune Laya, the open System One decision model, for any typed-decision…
wh000wh000/awesome-jev-live
Frequently asked questions
Why fine-tune Laya?
Out of the box it is a generalist with uncalibrated confidence and short context; fine-tuning on one task closes most of the gap with Jev.
Where does training data come from?
An LLM writes labelled cases and a second LLM filters wrong labels.
How is quality checked?
On a hand-written test set never used for training.
Related projects
CLI and Claude Code skill that turns one typed decision into a fine-tuned, calibrated, evaluated Laya checkpoint you can serve.
ModelbetaPythonComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython Joke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPython11 MB CPU-only model that answers choice, noul and score questions with fitted calibration and no LLM at inference.
ProjectresearchPython