{
  "seed": 1234,
  "total_samples": 588,
  "total_note": "512 chat-templated conversations + 76 raw-completion strict-format samples",
  "max_tokens_per_sample": 4096,
  "conversational_mix": {
    "ja_chat": {
      "count": 205,
      "target": 205,
      "sources": [
        "llm-jp/oasst2-33k-ja"
      ],
      "licenses": [
        "apache-2.0"
      ],
      "mean_tokens": 418.33170731707315
    },
    "ja_knowledge": {
      "count": 102,
      "target": 102,
      "sources": [
        "wikimedia/wikipedia:20231101.ja"
      ],
      "licenses": [
        "cc-by-sa-4.0"
      ],
      "mean_tokens": 1440.2941176470588
    },
    "ja_math": {
      "count": 77,
      "target": 77,
      "sources": [
        "Kendamarron/magpie-japanese-math-instruction-17k-qwen2.5-bakeneko-32b-instruct"
      ],
      "licenses": [
        "apache-2.0"
      ],
      "mean_tokens": 457.0
    },
    "code": {
      "count": 51,
      "target": 51,
      "sources": [
        "ronantakizawa/python-code-instructions-japanese"
      ],
      "licenses": [
        "mit"
      ],
      "mean_tokens": 243.6078431372549
    },
    "en_chat": {
      "count": 77,
      "target": 77,
      "sources": [
        "HuggingFaceH4/ultrachat_200k"
      ],
      "licenses": [
        "mit"
      ],
      "mean_tokens": 1404.7922077922078
    }
  },
  "strict_format_slice": {
    "count": 76,
    "form": "raw text in a zero-shot instruction scaffold (### 指示/### 回答形式/### 入力/### 応答), NOT chat-templated at PTQ time",
    "nli_answer_tag": {
      "count": 38,
      "source": "zenless-lab/jnli train split",
      "license": "cc-by-sa-4.0 (JGLUE)"
    },
    "qa_answer_tag": {
      "count": 22,
      "source": "wikimedia/wikipedia 20231101.ja, skip 12000 (disjoint from the perplexity held-out set and the conversational calibration set)",
      "license": "cc-by-sa-4.0"
    },
    "json_schema": {
      "count": 16,
      "source": "wikimedia/wikipedia 20231101.ja",
      "license": "cc-by-sa-4.0"
    },
    "rationale": "constrained-output formats (tags, JSON, fixed labels) are served as raw zero-shot prompts by many harnesses and applications, so they are calibrated in that form as well"
  },
  "base_model_revision": "098f2b2cf33021eba19a6d3582aa3d071ccc0aff",
  "calibration_file": "calib_full.jsonl (bundled, 588 lines)"
}