{
  "version": "0.1.0",
  "model_id": "Dohnuts-0.1.0-0.8B",
  "checkpoint_sha256": "196be33a0282537bcd821e2115643b352d0ad2a0bbaf7b242a1b7fe5bd96cfdf",
  "seed": 42,
  "step": 3600,
  "matplotlib": "3.11.2",
  "numpy": "2.5.3",
  "figures": [
    {
      "name": "overview",
      "title": "Model overview and benchmarks",
      "notes": [
        "Dohnuts-0.1.0-0.8B, seed 42, update 3600; one RX 7900 XTX.",
        "Benchmark values come from chart-data.csv; held-out macro accuracy comes from results/quality.csv.",
        "Panel captions identify evaluation coverage, reference exposure, precision and warm-cache conditions."
      ]
    },
    {
      "name": "01_jevbench",
      "title": "JevBench: decision types and difficulty",
      "notes": [
        "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."
      ]
    },
    {
      "name": "02_applications",
      "title": "Laya multilingual: application tasks",
      "notes": [
        "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."
      ]
    },
    {
      "name": "03_multilingual",
      "title": "Laya multilingual: language coverage and accuracy",
      "notes": [
        "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\u00d7 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."
      ]
    },
    {
      "name": "04_vision",
      "title": "Laya Vision: accuracy and calibration",
      "notes": [
        "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."
      ]
    },
    {
      "name": "05_latency",
      "title": "Same-hardware inference latency",
      "notes": [
        "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."
      ]
    },
    {
      "name": "06_agreement",
      "title": "JevBench: paired correctness and answer agreement",
      "notes": [
        "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."
      ]
    }
  ],
  "files": {
    "01_jevbench.svg": "b9ef1fd65c44bd3b216f2b583bcf00e3ae4b78e6d0bdf1a3dd19342e8dbfdfde",
    "02_applications.svg": "3f4e072de001da5c29af6895caaf59a86399e2dfc6d78bed23eb34cdc1d75034",
    "03_multilingual.svg": "efb052e46f2f0f1b050ab82ea0c2a5d22e12d6a83a8418d95ab1df0a6d9bcccd",
    "04_vision.svg": "4001f6d79b66283939deb0fdca09e39d899055187fd25495b5d2f2beefb883f7",
    "05_latency.svg": "4b515d7e439c6a87e069d088a86faf15746d8591d950116f4e9c7ede4154b7e1",
    "06_agreement.svg": "a4cfa64f1b211bf15f3326f8e82a5ff656409b191a0cc96e7b0b363765ce7822",
    "chart-data.csv": "3e912fc30bc4ae0a39c55c2191c79459629582f720ca9410b44ec88df50c0b54",
    "overview.svg": "91dcae40936da68953638c6ffcdfca750f718b6883fed0affba376cad8da8349"
  }
}