minojev

简体中文</a> · <b>English</b> · <a href="https://zeredy879.github.io/minojev/">Live demos</a> · <a href="https://zeredy879.github.io/minojev/benchmark.html">Benchmark</a> · <a href="https://huggingface.co/zeredy879/minojev">Models</a> · <a href="https://huggingface.co/datasets/zeredy879/minojev-data">Data</a>
</p>

<p align="center">
  <a href="https://huggingface.co/zeredy879/minojev"><img alt="Hugging Face models" src="https://img.shields.io/badge/Hugging%20Face-models-f0b06a"></a>
  <a href="https://huggingface.co/datasets/zeredy879/minojev-data"><img alt="Hugging Face datasets" src="https://img.shields.io/badge/Hugging%20Face-datasets-59d3a8"></a>
  <img alt="license" src="https://img.shields.io/badge/license-MIT-f0b06a">
  <img alt="decode steps" src="https://img.shields.io/badge/decode__steps-0-6ea8fe">
  <img alt="head params" src="https://img.shields.io/badge/decision__head-0.8M-8d9bb3">
</p>

> **The one-sentence version.** By the time a model finishes reading your question it
> already has an opinion — minojev reads that opinion back as a typed, calibrated
> distribution instead of making the model write a sentence.

## minojev's TLDR

- **Typed decisions, not text.** `choice` (2–255 candidates), `boolean`, and `score`
  (2–10 levels) return complete probability distributions.
- **Zero output tokens.** Every record reports `decode_steps: 0`; there is nothing to
  parse and no format to break.
- **Head training is the core.** Freeze any language model, train only the small decision
  head, and get a calibrated decision layer in minutes.
- **Measured against generation.** On the bundled 120-decision suite the head-trained
  model scores **95.8% vs 80.0%**, with 0 output tokens and a ~5× better p95 latency.
- **A zero-training route too.** The native-logits engine re