Does Luce generate text answers?

No. Inference returns a probability per option for Choice, Score and Noul questions; the README states that no text is generated at inference, which is what keeps latency low.

What data do I need to ship a task?

luce synth can generate inputs or annotate existing data with any OpenAI-compatible teacher. The README recommends 100-300 real labeled items so eval can fit temperature on them.

How do I install Luce?

Install from git with pip, since the PyPI release is pending; the base install is pip install git+https://github.com/scienthoon/luce, and adding the [synth] extra enables embedding dedup.

Which teacher can Luce use to generate data?

Any OpenAI-compatible endpoint. The teacher is never defaulted, so you pass --teacher URL|MODEL, and the key is read from LUCE_TEACHER_API_KEY, then OPENAI_API_KEY, then AI_GATEWAY_API_KEY; Ollama needs none.

What hardware does Luce training need?

The README trains Qwen3-4B-Base with a LoRA and decision head on a single RTX 4070 SUPER 12 GB, about 40 minutes per task over 2-3 epochs, and the browser demo replays recorded outputs so it needs no GPU.

Does Luce's calibration transfer to unseen tasks?

No. Accuracy transfers better than calibration: on datasets Luce never trained on, the README measures a refit temperature of 2.5 to 3.3, meaning the model is over-confident on unseen tasks.

What happens to low-confidence Luce answers?

Serving with --review queues inputs whose answers fall under 0.9 confidence for human review; the demo shows them dropping into a review queue while high-confidence tickets are answered automatically.

CommunityJev Models & Reimplementations

Luce

Open recipe for calibrated decision models: an LLM teacher writes training data from a one-sentence task description, a LoRA plus decision head on Qwen3-4B-Base answers Choice, Score and Noul questions with calibrated probabilities, evaluated next to Jev on identical items.

See highlights

Open github.com

Stars: 8
Language: Python
License: APACHE-2.0
Published by: Community
Last verified: Oct 2026

scienthoon/luce

User repository on GitHub

View on GitHub

Luce: a recipe for calibrated decision models — a sentence about your task in, a small model that answers typed questions with honest probabilities out (init → synth → train → eval → serve)

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Describe the decision. Supply your inputs or generate them with an LLM. Luce trains a LoRA + decision head on an open model and serves decision probabilities.

Quoted from the project's README. Read the full README.

Tagscommunity calibration training lora local

Works forJev for Classification Jev for Scoring Jev for Evaluation

Used inJev for Customer Support

Languages: Python
Capabilities: Open sourceSelf-hostedRuns locallyRequires API key

Highlights

Five CLI stages: init scaffolds questions, synth generates or labels data, train fits the head, eval reports accuracy and ECE, serve hosts POST /v1/ask
Trains on one RTX 4070 SUPER 12 GB in about 40 minutes per task using Qwen3-4B-Base and 2-3 epochs
On 2,964 rule-generated support tickets it reports 91.1% combined accuracy with ECE 0.022, against Jev zero-shot at 75.1
Teacher is any OpenAI-compatible endpoint (Vercel AI Gateway, OpenAI, Ollama); outputs are probabilities and no text is generated at inference
Browser replay demo on Hugging Face Spaces runs without a GPU; low-confidence inputs queue for review

Watch out

Apache-2.0. The PyPI release is pending so install from git; calibration does not transfer to unseen tasks, and the README documents maze experiments where training added nothing over a constant predictor.

Questions about this entry

Short answers drawn from the linked project or page.

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Laya Typed-Decisions

The fine-tuned Laya checkpoint for four typed-decision workflows: 0.766 accuracy on 2,000 held-out decisions, above Jev 1.13.0's published 0.727 and the teacher self-agreement ceiling it trained past.

Jev Models & Reimplementations#community#open-models#local

Community

Gutsy

An Apache-2.0 0.8B decision model that returns calibrated probabilities for yes/no, choice and score questions on CPU. Distributed as GGUF for llama.cpp and served behind a Jev-style /v1/systemone API, so existing Jev clients mostly need only a new base URL.

Jev Models & Reimplementations#community#gguf#local-inference

Community

kouhxp/gutsyHTML

WaterSheep

Apache-2.0 ModernBERT fine-tune that answers noul, choice, score and multi-label questions with a probability for every option. It serves Jev's POST /v1/systemone locally so TypeSafe's Python SDK runs unchanged, reporting 0.026 expected calibration error on its test split.

Jev Models & Reimplementations#community#calibration#modernbert

Community

SamratDuttaOfficial/WaterSheepPython

Decider Chat Gemma-4-31B

An untrained readout: stock Gemma-4-31B weights plus a decider config that reads option letters at the answer slot, scored #2 of 70 on Decision Index v0.2.1 at 57.33 against Jev 1.13.0's 57.91.

Jev Models & Reimplementations#community#open-models#local

Community

Rizzo Flow

877

A local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.

Jev Models & Reimplementations#community#python#open-models

Community

Rizzo-AI-Academy/rizzo-flowPython

AutoJev

130

An open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.

Jev Models & Reimplementations#community#calibration#qwen

Community

denis-pplx/autojevPython

From the Jev guides

Original write-ups on what this entry is for.

Multimodal Decision Models: What Clef and d1 Change

Oct 12, 2026

Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.

Jev Models & Reimplementations#multimodal#open-models#decision-model

Read article

The Decision-Model Price War

Oct 11, 2026

One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.

Jev Models & Reimplementations#decision-model#benchmarks#pricing

Read article

What Is Actually on Hugging Face

Oct 8, 2026

Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.

Jev Models & Reimplementations#open-models#hugging-face#local

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai