知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

Calibration & Research

Use this category for work that studies or exploits Jev's calibrated confidence — RLCD-style training, probability quality, threshold selection, uncertainty analysis.

Submission format

- [Name](URL) - Industry: one-sentence description of the Jev use case.

Entries

decider - Open models: reproduces the System One shape with a Qwen3.5-2B fine-tune that emits typed decisions with calibrated probabilities in one pass.
openjev - Open research: independent local preview that answers bilingual probability questions from context, questions, and candidate answers, inspired by TypeSafe Jev.
Parallel Constrained Decoding (Qwen2.5-1B-RLCD) - Open research: RLCD-trained Qwen2.5-1B demo exploring open-source parallel constrained decoding as an alternative to Jev.
NanoJev - Open replica: a 0.6B parallel decision model that returns full probability distributions with no output-token decoding, shipped with its training pipeline, weights, and dataset.
open-alternative-jev - Open alternative: runs a Jev-shaped decision model locally on your own GPU.
mini-jev - Local reproduction: implements Jev's typed-decision interface on top of a local LLM.
Laya - Open alternative: non-autoregressive decision model that answers choice, score, and noul questions with RLCD-trained calibrated probabilities in a single ~35 ms forward pass, published on PyPI and Hugging Face.
Jev-compatible public API - Open research: a public Jev-shaped API backed by an open Qwen3.6-35B-A3B model so anyone can try the typed-decision interface.
kev - Trainable replica: a tiny Jev-like model on top of Qwen2.5-0.5B that trains and runs on a MacBook, shipped with its own research runs and evaluation scripts.
jevinci - Creative experiment: paints images by having Jev predict every pixel's colour in parallel, with predicted confidence deciding how wide each stroke is drawn.
jev-local - Local reproduction: Jev-compatible POST /v1/systemone server answering typed Choice/Score/Noul questions with confidence from open weights, verified as an official-SDK drop-in with temperature-fit calibration (set3 n=1316, 0.83 overall).
LitJev - Local reproduction: a reproduction of Jev that turns any Qwen model into a fast decision model, serving the same /v1/systemone schema (Choice, Score, Noul) with no training and no generated answer text.
CUA-S1-FORMS - Specialist decision model: a 706,048-parameter, 2.8 MB jev-like option scorer that rates FILL / CHECK / CLICK / SKIP for each form field in one parallel pass, reporting 99.7% on its own form-filling eval against Jev's 83.6% - a specialist on home turf rather than a general win.
jevlike - Training library: build a small model that chooses among a changing list of text options and returns one probability per option in a single pass - the base CUA-S1-FORMS was built on.
jevbetter - Improved scorer: a stronger one-pass scorer over a variable list of text options, using a hashed n-gram encoder, rival-aware attention, and gated heads.
jevlike-esp32 - Edge deployment: exports a jevlike scorer as ESP32 firmware with a C scorer and a host-side check, putting one-pass decisions on a microcontroller.
von - Open alternative: a 395M non-autoregressive System One model that answers typed questions with calibrated probabilities in under 15 ms, positioned as a local drop-in replacement for Jev.
JevForge - Open research: an end-to-end stack for auditable data construction, Qwen3.5-0.8B training, fixed Mind2Web and OOD evaluation, local serving, and a preliminary RLCD baseline.
minojev - 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).
Luce - Open recipe: describe the decision task in a sentence, an LLM teacher writes the training data, a LoRA + decision head on Qwen3-4B-Base answers choice/score/boolean questions with calibrated probabilities in one forward pass; trains on a 12 GB card. Reports accuracy and ECE next to Jev on identical test items (rule-generated tickets 91.1 vs 75.1, phishing 97.4 vs 62.6, GitHub issue priority 41.1 vs 37.5); browser replay demo without a GPU.
poorjev - Local reproduction: implements Jev's typed Choice/Score/Noul interface on commodity zero-shot NLI models and makes the confidence honest with temperature scaling and conformal abstention, shipping a reproducible calibration eval (ECE 0.170 to 0.071, cross-validated) that runs offline with no API key.
openJev-verdict-2.0 - Open decision engine: a calibrated 151M non-autoregressive model that reports beating both TypeSafe Jev and Laya on typed-decision benchmarks, shipped with its own test suite.