Open-Jev-9B
A LoRA adapter and decision head on Qwen/Qwen3.5-9B, by zefan-cai, read from the Hugging Face
model card and cross-checked against the GitHub repo and the training dataset's own card.
Curator-sent, "ingest this," 2026-09-24. Like laya, it returns typed decisions — Choice,
Noul, Score — without generating text, and its own project README states the target directly:
"an independent implementation inspired by TypeSafe's Jev."
What it is
Rank-8/alpha-16 LoRA over a 9B autoregressive base, plus a scalar decision head initialised from
the base model's own Yes-minus-No logit readout and fine-tuned jointly with the adapter. That
initialisation is the architectural fingerprint: rather than training a decision head from scratch
on top of a frozen encoder (laya's approach), Open-Jev-9B starts from a signal already latent in
a generative chat model — whether it says "yes" or "no" — and adapts it. The HF tag list calls the
result non-generative: at inference it scores candidates and returns probabilities, never sampling
a token, even though the backbone underneath is a text generator.
Training (9B checkpoint): one epoch over the release-v2 split of the Open-Jev
dataset — 80,816 rows (79,116 redistributable, the rest withheld) — at global batch 4, 20,204
optimiser steps. Temperature 1.897, calibrated on a separate 512-row slice. The card does not name a
loss function beyond "trained jointly with LoRA"; the project's own README documents a 27B v1.1
sibling trained with "soft cross-entropy + 0.1 Brier loss" on a larger 328,672-row mixture — a
supervised, calibration-shaped loss, not the policy-gradient RL either Jev
or laya use. Whether the 9B checkpoint shares that recipe isn't stated on its own card. If it
does, Open-Jev-9B reaches the same typed-decision interface and the same calibration target as the
RLCD family without RLCD's REINFORCE machinery — evidence that calibration doesn't strictly
require policy-gradient RL, not just a cheaper imitation of a technique that does.
Evaluation, on its own held-out split
| Split | Rows | Hard accuracy | Expected accuracy | NLL | Brier | ECE |
|---|---|---|---|---|---|---|
| Test | 10,532 | 97.54% | 94.72% | 0.131 | 0.039 | 0.008 |
| OOD | 15,920 | 91.97% | 90.40% | 0.299 | 0.127 | 0.037 |
These numbers are not comparable to Laya's 0.766 typed-decisions accuracy, and the card is explicit
about why not to over-read them. Three separate reasons, not one: the eval set is Open-Jev's own
release-v2 held-out rows, drawn from the same mostly-synthetic, controlled generator that made
the training data — an easier bar than scoring against an external benchmark or a competitor's eval
set, the way Laya's card scores against Jev's. The dataset's own card qualifies "OOD" further: it
means source-specific variation within the generator's own configs, not held-out real-world
domains. And the model card states its own limit in plain terms: "synthetic held-out decision
metrics do not establish broad real-world reliability, closed-loop browser/game/flight success, or a
calibrated probability guarantee outside the evaluated distribution." A vendor saying that about its
own numbers is unusual, and it's the reason this page treats 97.54%/91.97% as descriptive, not as a
result that outranks Laya's or Jev's.
Independence, stated where Laya's card is silent
This closes a question laya's page left open. Open-Jev's README: "Open-Jev is an independent
implementation inspired by TypeSafe's Jev. It does not reproduce proprietary RLCD, private weights
or training data." The dataset card repeats it: "not affiliated with TypeSafe." Laya never states
whether it drew on Jev's design beyond the benchmark comparison table; Open-Jev names the relationship
outright and disclaims reproducing the parts it can't verify (RLCD's internals, Jev's weights, Jev's
data). Three projects now share the choice/score/noul vocabulary — Jev coined it, Laya reused it
without comment, Open-Jev reused it with an explicit citation.
Scale and licence
Sits alongside 2B and 27B v1.1 siblings in the same project (not read in detail here — this page covers only the 9B checkpoint's own card). Apache-2.0 for the adapter and head; MIT for the training and inference code; CC0-1.0 for the dataset's own generated content (explicitly not covering upstream wording, external assets or source data folded into it). GitHub repo created 2026-09-20, 294 stars / 28 forks as of this ingest four days later — ordinary growth, unlike laya's.
Where it sits
Runner-up spoke: llm-providers-wiki, same reasoning as laya — system-one-models-and-jev owns Jev's market position; this page is the model itself. Both Laya and Open-Jev-9B are read here as two independent, unrelated attempts at the same target Jev defined, arriving within days of each other and of this wiki's own ingest of the first one.
Tier T3: architecture and licence facts are checkable artifact facts; every accuracy, calibration and Brier number is the maker's own measurement on the maker's own held-out data, which is what T3 means, and the card's own limitations section says so before this page has to.