Jev Colab LabReproducible decision-model experiments

Five pinned GPU experiments, sanitized evidence, and practical Colab reproduction paths.

Start with the guide

日本語

Browse the source

🎯

Pinned sources

Every experiment records upstream repositories, model revisions, licenses, and intended GPU target.

📏

Separated measurements

Loading, first inference, warmup, steady state, and peak VRAM stay distinct where the runner supports them.

🧼

Sanitized artifacts

Results exclude credentials, session metadata, private inputs, and model weights.

Snapshot

3 × T4

Laya, Kev-0.5B, and Jevlike short training/inference.

2 × L4

SemIf direct option readout and OpenJev NLI.

5 result sets

Each result is linked to a checked-in JSON record and experiment README.

5 JevDash captures

Independent single-episode observations; not a completed ability comparison.

The five experiments are intentionally heterogeneous. This site is a reproducibility index, not a leaderboard: fixtures, readouts, model sizes, and probability semantics differ.

Continue

Getting started — local checks, uv, WSL, and Colab boundaries.
Experiments — what each runner measures and where its evidence lives.
Reproducibility — session isolation, sanitized outputs, and result interpretation.
Sources and licenses — upstream revision and license map.

Current measurement note

Jevlike timing fields keep their measurement boundary in the result schema and are not used as cross-experiment comparison points.