CalibratedDecisions.

Repo · Benchmarks & research

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…

Open github.com ↗

From the repository

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)

How builders describe it

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.

The decision Jev makes

Benchmark questions with known answers, to check accuracy and confidence.

Where it fits

Head-to-head tests, calibration studies and independent research into how well Jev decides, how fast, and at what cost. All 682 benchmarks & research projects →

Related projects