LLM2Jev: Turn LLMs into Jev-Style Decision Models

简体中文

LLM2Jev adapts local language models to Jev-style structured decisions. It accepts runtime-defined Choice, Score, and Noul questions and returns typed answers with probabilities.

LLM2Jev is an independent open-source project. It is not affiliated with or endorsed by Jev or TypeSafe.

Quick Start

On Linux with a supported NVIDIA GPU, run a local model through SGLang:

git clone https://github.com/Yinsongxu/LLM2Jev.git
cd LLM2Jev
uv sync --extra sglang
source .venv/bin/activate
python examples/sglang_inference.py --model-path /path/to/model

The example submits Choice, Score, and Noul questions and prints the response as JSON. Replace /path/to/model with a local Hugging Face-compatible causal language model directory.

Key Features

• Structured decisions: define Choice, Score, and Noul questions at runtime. Get option probabilities, weighted scores, or the probability that a condition is true; Choice and Score also include confidence.
• Probabilities from logits: score each candidate with an independent yes/no judgment, then assemble JSON in code. No answer tokens are generated.
• Shared-prefix caching: stage candidate submissions to reuse SGLang's Radix Cache within a single request, including a first request with no relevant cached prefix.

Candidates share state, and candidates for the same question also share its instructions. LLM2Jev first scores a real criteria candidate to establish the prefix cache, then submits candidates that can reuse it. Each candidate is scored once, reducing repeated computation for long inputs with many candidates.

Learn how it works: From Jev Request to LLM Request → Shared-prefix design.

Installation

See Installation for environment requirements, SGLang and Transformers dependencies, and uv or pip installation.

Getting Started

See the Usage guide for complete examples:

• SGLang Python API
• Transformers backend
• System One HTTP API
• Choosing between staged and all

Benchmarks

See Performance benchmarks for the Qwen3-1.7B / RTX 5090 measurements, test conditions, and comparison of staged and all across cold and warm caches. Gains depend on input length, candidate count, and cache state.

Roadmap

• More benchmarks across model sizes, datasets, and workloads, covering decision quality, latency, and throughput.
• An interactive web demo for submitting questions and inspecting probabilities.
• Multimodal model and input support.
• More multimodal tasks and demos.

Tests

python -m unittest discover -s tests -v

License

This project is licensed under the Apache License 2.0.