Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

openjev-experiments

Single-script CLI and library for NLI classification and multiple-choice reranking with the AlexWortega/openjev Qwen3.5-4B cross-encoder.

Projectprototype
View on GitHub ↗

Overview

openjev-experiments wraps AlexWortega/openjev, a Qwen3.5-4B cross-encoder that reads a premise and a hypothesis and answers contradiction, entailment or neutral. A self-contained classify.py offers pair classification, multiple-choice reranking (pick the option best entailed by a question), a library API and a CLI, with batch input from JSONL or CSV.

The README shows an example pair returning entailment at 0.8727 and a rerank picking carbon dioxide for a photosynthesis question. It documents device and precision options, batch size, truncation length and CPU thread handling for machines with limited RAM. The checkpoint is about 9.1 GB. It suits developers experimenting with option-logit style decisions on local hardware.

What the decision model decides

TypeQuestionAnswers
ChoiceDoes the premise entail, contradict or stay neutral on the hypothesis?entailment · contradiction · neutral
ChoiceWhich option is best entailed by the question?—

How it works

  1. Install torch and transformers; download the checkpoint.
  2. Run classify for premise/hypothesis pairs or rerank for options.
  3. Read class probabilities or the winning index.
  4. Batch from JSONL or CSV.

Strengths

  • Single self-contained script.
  • Runs on CPU, MPS or CUDA.
  • JSON output for scripting.

Caveats

  • 9.1 GB checkpoint.
  • NLI framing only; not a general System One server.

Getting started

.venv/bin/python classify.py classify --premise "A man is playing a guitar." --hypothesis "Someone is making music."

At a glance

Type
Project
Status
prototype
Language
Python
Runtime
CLI, Transformers
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What model does it use?

AlexWortega/openjev, a Qwen3.5-4B NLI cross-encoder.

Can it pick between options?

Yes, the rerank command returns the option best entailed by the question.

How much disk does it need?

About 9.1 GB for the checkpoint.

Related projects

  • go-system-one

    Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.

    ProjectbetaGo
  • hearim

    Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.

    ServicebetaGo
  • ifReflex.cpp

    llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.

    ProjectbetaC++
  • llm2jev

    Turns any chat model on SGLang, vLLM, transformers or MLX into a Jev-compatible service with one prefill and one label token per question.

    ProjectbetaPython
  • verdict

    Self-hosted Jev-compatible System One server that sits in front of llama-server and reads option probabilities from any GGUF model.

    ServicebetaPython
  • DuoMind

    OpenAI-compatible local server pairing a small llama.cpp model for generation with Jev for the closed decisions.

    ServiceprototypePython