minojev
Jev-style replica: frozen Qwen3-1.7B plus a ~0.8M-parameter trained head returns calibrated Choice, Boolean and Score distributions.
Overview
minojev is an independent, reproducible Jev-style decision model. Its core idea is head training: freeze a language model backbone (Qwen3-1.7B in the release), cache the final hidden state for each state-question-option path once, and train only a small decision head of about 0.8M parameters, then fit one calibration temperature per primitive on a dev split. It supports choice (2-255 candidates), boolean and score (2-10 levels), with zero output tokens.
On its bundled 120-decision suite (banking77, CLINC150, Amazon Polarity, GSM8K verification), the README reports 95.8% accuracy and ECE 0.024 for the head-trained model, versus 88.3% for a zero-training native-logits readout and 80.0% for a same-backbone generative baseline, with better p95 latency. At confidence of at least 0.9 it reports 89.5% coverage at 98.9% accuracy on its full test suite. List summaries also mention a 547k-parameter from-scratch variant trained on CPU. It is aimed at researchers and builders who want a transparent, laptop-trainable decision layer.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of 2-255 runtime-defined candidates fits the state? | — |
| Yes / no | Boolean question about the state | true · false |
| Score | Score on a 2-10 level scale | — |
How it works
- Freeze the backbone and cache candidate features once.
- Train the small decision head with distribution losses.
- Fit a temperature per primitive on held-out data.
- Serve decisions and accept above a confidence threshold, escalating the rest.
Strengths
- Zero output tokens, so nothing to parse and no format failures.
- Training the head takes minutes on a laptop per the README.
- Calibrated confidence suitable for accept/escalate policies.
- Committed datasets, predictions and ECE results for reproducibility.
Caveats
- The README notes an out-of-distribution boundary on never-trained workloads.
- Headline numbers come from the project's own balanced suites.
At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- service
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Open replica: a 547k-parameter model that answers runtime-defined `Choice` (2-255 candidates), `Boolean`, and `Score` questions with dev-calibrated distributions in one forward pass and zero output tokens, trained from scratch on CPU with committed datasets, predictions, and ECE results (maze 0.016).
yibie/awesome-jevDecisions, not tokens: minojev reads calibrated, typed probability distributions straight from hidden states in one forward pass — zero output tokens, fully reproducible on a laptop CPU.
logicrw/awesome-jev-projectsHead-training reproduction: a frozen Qwen3-1.7B backbone plus a ~0.8M-parameter decision head returns calibrated Choice/Boolean/Score distributions in one forward pass with zero decoded tokens; ships an 8k-request converted dataset, source-isolated OOD evaluation, and a same-backbone generation baseline (95.8% vs 80.0% on its own balanced suite, ECE 0.024).
cobanov/awesome-jevIndependent Jev-style replica with head training: a frozen Qwen3-1.7B backbone plus a small trained decision head returns calibrated Choice/Boolean/Score distributions in one forward pass, with committed datasets and a same-backbone generation benchmark.
hellogumbo/awesome-jev
Frequently asked questions
What is head training here?
The language model stays frozen and only a ~0.8M-parameter head is trained on cached hidden states, then calibrated with a temperature.
How does it compare to prompting the same model?
On its 120-decision suite the README reports 95.8% versus 80.0% for the same backbone generating a single-letter answer.
Do I need a GPU?
The README says the release was trained in minutes on an Apple Silicon laptop; a smaller variant trains on CPU.
Related projects
Exact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPythonOpen decision models that derive calibrated confidence from recurrent looped-layer dynamics instead of RL, with early exit and a Jev API.
ProjectresearchPythonLaptop-scale reimplementation of RL for Calibrated Decisions: parallel constrained decoding trained with GRPO and a Brier reward.
ProjectresearchPython- AnyJev★ 1,255
Nokia's training-free library that turns open LLMs into Jev-style deciders robust to option order, plus self-distilled Tacit models.
ModelresearchPython