Top score in view
—
—
Matched model evaluation
Same test. Different models.
Compare Jolt, Dohnuts, Laya, and Kev on the same frozen examples. Start with the shared text test, then switch to image results or the full 26-dataset view where coverage allows.
Built October 17, 2026 · Works offline
How to read this report. “Dohnuts 0.8B (retrained)” is the report name for Jolt’s B0 control. It was trained from the pinned Qwen base on Jolt’s training split with the Dohnuts recipe. “Dohnuts 0.8B (released)” is the separately trained downloaded checkpoint. The primary score is dataset-macro accuracy. Every dataset gets equal weight, so large image datasets do not drown out smaller tasks. Use row-weighted accuracy to see the result when every example counts equally. Results from different test suites appear below as diagnostics, not as part of the shared-test ranking.
Separate benchmark comparison
JevBench · public tasks
Headline score is the mean of easy, standard, and hard accuracy, with each tier weighted equally. Pooled accuracy is shown beside it.
231 public tasks · 303 official tasks unavailable
This is a separate suite, so these scores do not enter the 26-dataset ranking above. Jolt and Dohnuts used a 4,096-token input limit; Laya used native 1,024-token truncation; Kev used its serving context. The table names each evaluation path.
Lead over next model
—
—
Shared evaluation data
—
—
Evidence per score
—
—
Aggregate ranking
Seed means are shown for Jolt runs; single-run results are labeled.
Whiskers show the minimum and maximum among Jolt seeds. They describe run-to-run variation, not a confidence interval. Scores for single-run models have no whisker.
Dataset detail
Select one or more models to compare their accuracy on each dataset in this lane.
Choose models
Scores by dataset
Cells show hard top-1 accuracy. Select a column heading to sort. The table follows the lane and model-family filters above.
What the comparison says
Training-data exposure and limits
Scoring every model on the same test rows makes the inputs comparable. It does not make their training histories equally clean.
Other evaluations
These results use other splits or task suites. They help describe each model, but they do not belong in the matched test ranking.
Jolt training experiments · development screens
The first chart shows capped development results at updates 400 and 800. Full-development scores appear only when that evaluation was run. These are training-selection measurements, not locked-test scores.
Kev official development suites
Kev's own development partitions, evaluated with its native benchmark code. These suites differ from the Jolt frozen test.
Jolt 0.8B L2 secondary suites
The additional application collection has its own task mix. Read their scores within each suite only.