Dohnuts-0.1.0-0.8B comparisons

Seed 42, update 3,600. Figures use saved evaluation results with fixed Laya multilingual and Laya Vision references.

Model overview · Chart data · Model identity and checksums

The figures are available as SVG. Captions document each evaluation protocol. One seed does not provide confidence intervals across seeds.

JevBench: decision types and difficulty

JevBench: decision types and difficulty

  • Jev: published per-task outcomes, v1.2.2. Dohnuts and both Laya checkpoints: local measurements. All 231 local outputs are valid.

  • Dohnuts context limit: 4,096 tokens. Laya defaults: 1,024 tokens with native truncation. Vision is tested on text here.

  • 303 official tasks are unavailable, including the judge tier. This is not the 534-task leaderboard score. One Dohnuts seed.

Laya multilingual: application tasks

Laya multilingual: application tasks

  • Counts beside tasks are decisions, not training examples. Original upstream question builders, seed 13 and candidate order.

  • The published application/chart reference is used consistently, including typed decisions (2,000 decisions).

  • Historical Laya dataset revisions and per-example hashes are unavailable; byte-identical historical inputs cannot be verified.

Laya multilingual: language coverage and accuracy

Laya multilingual: language coverage and accuracy

  • MASSIVE intent uses 20 candidates; XNLI uses 3. The 51-language and 14-language suites are distinct evaluations.

  • On the 51-language suite, accuracy exceeds 15% (3× random) in 51/51 languages for Dohnuts and 45/51 for multilingual.

  • Dohnuts: measured. Laya multilingual: published reference. Historical byte identity cannot be independently verified.

Laya Vision: accuracy and calibration

Laya Vision: accuracy and calibration

  • ScienceQA: official test image subset. VQAv2: independent split of the validation pool. These are not the model-card splits.

  • Laya Vision has possible VQAv2 training exposure and A-OKVQA checkpoint-selection exposure; this is not a blind comparison.

  • Laya: FP32 CPU, 2,048-token total/head budgets. Dohnuts: BF16. ECE: 15 bins, deployed temperatures, no per-task refit.

Same-hardware inference latency

Same-hardware inference latency

  • 3 warmups and 20 synchronized repetitions. Includes preprocessing and transfers; excludes loading, network and queueing.

  • Uses each native API, tokenizer and cache behavior. Image timings are warm, not uncached image-encoder timings.

  • Laya uses the fixed same-hardware baseline; Dohnuts uses Dohnuts-0.1.0-0.8B. Jev API latency is not mixed into this chart.

JevBench: paired correctness and answer agreement

JevBench: paired correctness and answer agreement

  • Jev publishes correctness, not predicted labels: its exact answer-agreement count is unavailable.

  • 148 shared correct answers confirm the same label. The 27 shared errors may have different labels; 175 is correctness agreement.

Reproduce the figures

Follow the plotting setup, then run from the repository root:

pdm run python scripts/plot_model_card.py --run runs/v1

The command writes PNG, SVG, and PDF figures to the experiment output directory. The published manifest identifies the model and checksums the figures and chart values.