What are the three lanes in this Space?

A cached plus parallel lane that prefills once and scores all branches in one batched forward, a naive lane that re-encodes the state per option as a live equivalence check, and a Qwen3.5-4B instruct baseline that generates JSON with no probabilities.

How does the cached path save work?

It prefills the state once and scores every question-option branch from that cache in one batched forward: two passes plus one per 96 branches beyond the first 96, versus one forward per question on the naive path.

Why is branching exact for Qwen3.5?

Qwen3.5 is a hybrid, so branching copies both the KV cache of its full-attention layers and the conv and recurrent state of its gated-delta-net layers; on random fp32 weights the branched scores match a full re-encode to about 1e-7.

What are the documented limits?

Training used 384-token sequences and at most 16 options; the Space accepts 16,384-token states but flags them out of distribution, refuses questions over 96 tokens and options over 64, and skips the instruct lane past 20,000 prompt tokens.

How are long document states handled?

Each branch carries its own copy of the state cache at roughly 32 KB per token, so long states get fewer branches per forward: at most 96, or 200,000 divided by the state token count.

Is the model commercially usable?

No. The ticket component of the training data is CC-BY-NC-4.0, so the card says the model is non-commercial.

What does the Load PR button fetch?

It uses the unauthenticated GitHub API, unless a GITHUB_TOKEN secret is set, to pull a public PR's title, description, file list, and per-file patches, leaving out lockfile diffs, labels, and changesets that state the answers.

CommunityJev Demos & Experiments

System One Scorer Demo

A three-lane demo of pngwn's System One 4B scorer: one prefill plus a batched forward for every option branch, a naive per-option equivalence check, and a Qwen3.5-4B instruct baseline generating JSON.

See highlights

Open huggingface.co

Format: Playground
Published by: Community
Author: pngwn
Last verified: Oct 2026

Tagscommunity playground demo gradio inference

Works forJev for Classification Jev for Scoring

Used inJev for Software & DevTools

Languages: Python
Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

Cached lane prefills once and scores every option branch in one batched forward: 2 passes plus one per 96 branches beyond the first 96.
Measures the scorer at 0.740 accuracy against a 0.508 majority baseline on 3,219 decisions about 500 real Gradio PRs, with ECE 0.047.
The untouched base model, read through letter logits, is as good on judgement questions and better above 2,000 tokens, the card reports.
Training sequences were truncated at 384 tokens and options capped at 16, but the Space accepts 16,384-token states and flags them as out of distribution.
The ticket component of the training data is CC-BY-NC-4.0, so the card states the model is non-commercial.

Watch out

The weights are pngwn/system-one-qwen3.5-4b-scorer, whose ticket training data is CC-BY-NC-4.0, so the model is non-commercial. The card frames the 384-token training length against 16,384-token inputs as an experiment, not a supported range, and long states allow only 200,000/state_tokens branches; wall-clock numbers come from shared ZeroGPU.

Questions about this entry

Short answers drawn from the linked project or page.

Related terms

Glossary definitions related to this entry.

Playground

More Jev resources like this

JEV-9B Decision Demo

An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.