JevForge

# JevForge

**An end-to-end toolkit for synthesizing decision data, training calibrated
candidate scorers, evaluating them, and serving Jev-compatible inference for
interactive web decisions.**

Data synthesis · training + calibration · evaluation · Jev-compatible serving

[Interactive demo](https://jev-forge.vercel.app) · [HF model](https://huggingface.co/AndeyTait/JevForge-0.8B) · [HF dataset](https://huggingface.co/datasets/AndeyTait/JevForge-Mind2Web) · [Dataset format](docs/DATASET.md) · [Chinese](README.zh.md) · [Design](docs/DESIGN.md) · [RLCD results](docs/RLCD_RESULTS.md)

</div>

JevForge turns structured decisions into one reproducible pipeline. Give it a
state, a question, and a candidate set; it builds training records, trains and
calibrates a scorer, evaluates fixed splits, and serves a complete probability
distribution through a Jev-compatible API. `choice`, `noul`, and ordered
`score` questions all use the same decision path.

Its first proving ground is the interaction loop at the heart of Jev-style
systems: choosing the next click, navigation target, form control, route, or
escalation from the elements currently available on a page.

<p align="center">
  <a href="https://jev-forge.vercel.app">
    <img src="./docs/assets/jevforge-demo.gif" width="960" alt="JevForge interactive web decision demo">
  </a>
</p>

## Results

Results use fixed, website-disjoint test and OOD splits. The 0.8B and 0.6B
models share the same data and evaluation protocol; Jev-1.13 is shown as a
separate reference arm.

| Metric | Raw Qwen3.5-0.8B¹ | JevForge 0.8B | JevForge 0.6B | Jev-1.13 |
|---|---:|---:|---:|---:|
| Test choice top-1 | 0.235 | **0.579** | 0.439 | 0.543 |
| OOD choice top-1 | 0.340 | **0.637** | 0.500 | 0.610 |
| Test noul accuracy / Brier | — | 0.826 / **0.12