v1.0.0/schema.json

{
  "schema_version": "1.0",
  "description": "Unified SFT reference dataset schema. Every row is a single training example with domain/subject search keys and a tiered veracity score.",
  "columns": [
    {
      "name": "id",
      "type": "string",
      "nullable": false,
      "unique": true,
      "description": "Globally unique row id. Scheme: {source_dataset with '/' replaced by '__'}--{row_index:07d}. Deterministic and traceable back to the source row."
    },
    {
      "name": "source_dataset",
      "type": "string",
      "nullable": false,
      "description": "Hugging Face repo id of the source dataset (e.g. 'databricks/databricks-dolly-15k')."
    },
    {
      "name": "domain",
      "type": "string",
      "nullable": false,
      "enum": [
        "science",
        "medicine",
        "mathematics",
        "computer_science",
        "history",
        "law",
        "finance",
        "literature",
        "general_knowledge",
        "commonsense_reasoning",
        "news",
        "government",
        "conversation",
        "education"
      ],
      "description": "Coarse subject area. One of the 14 allowed values."
    },
    {
      "name": "subject",
      "type": "string",
      "nullable": false,
      "description": "Finer-grained topic label (free text), e.g. 'grade-school math word problems', 'US government reports', 'competitive programming'. Used for subject-level search/filtering."
    },
    {
      "name": "format",
      "type": "string",
      "nullable": false,
      "enum": ["instruction_response", "raw_text"],
      "description": "Row format. instruction_response rows populate instruction/response; raw_text rows populate text."
    },
    {
      "name": "instruction",
      "type": "string",
      "nullable": true,
      "description": "The prompt/instruction/question. Required when format == instruction_response; null when format == raw_text."
    },
    {
      "name": "response",
      "type": "string",
      "nullable": true,
      "description": "The target response/answer. Required when format == instruction_response; null when format == raw_text."
    },
    {
      "name": "text",
      "type": "string",
      "nullable": true,
      "description": "Raw text body. Required when format == raw_text; null when format == instruction_response."
    },
    {
      "name": "veracity",
      "type": "integer",
      "nullable": false,
      "minimum": 1,
      "maximum": 5,
      "description": "Tiered provenance/accuracy score. 5 = highest confidence (peer-reviewed/official/authoritative), 1 = unknown/unverified. See design.md rubric."
    },
    {
      "name": "veracity_basis",
      "type": "string",
      "nullable": false,
      "description": "JSON string recording the evidence for the veracity score. Shape defined below."
    },
    {
      "name": "license",
      "type": "string",
      "nullable": false,
      "description": "License string for the source dataset. Use the exact card/metadata value where available; otherwise 'undeclared' plus the original-source license in parentheses where known."
    }
  ],
  "veracity_basis_shape": {
    "description": "Exact JSON object serialized into the veracity_basis column.",
    "fields": {
      "source": "string — HF repo id (same as source_dataset)",
      "license": "string — license string",
      "provenance": "string — one of: human-written, human-curated, synthetic, web-crawled, peer-reviewed, official, mixed",
      "tier": "integer 1-5 — the veracity tier",
      "spot_checks": "array of objects — one per sampled row actually inspected by the build thread",
      "spot_check_summary": "string — 1-2 sentence summary of spot-check outcome (e.g. '20/20 samples matched gold answer key')"
    },
    "spot_check_item_shape": {
      "sample_id": "string — id of the sampled row",
      "checked": "boolean — true if the sample was actually inspected",
      "result": "string — one of: pass, fail, flag",
      "note": "string — what was checked and what was found"
    },
    "example": {
      "source": "openai/gsm8k",
      "license": "MIT",
      "provenance": "human-written",
      "tier": 5,
      "spot_checks": [
        {"sample_id": "openai__gsm8k--0000001", "checked": true, "result": "pass", "note": "final answer 72 matches step-by-step solution"}
      ],
      "spot_check_summary": "20/20 sampled solutions recomputed to the stated final answer"
    }
  },
  "invariants": [
    "format == 'instruction_response' => instruction is non-null AND response is non-null AND text is null",
    "format == 'raw_text' => text is non-null AND instruction is null AND response is null",
    "veracity is an integer in [1,5]",
    "domain is one of the 14 enum values",
    "id is unique across the whole dataset",
    "veracity_basis parses as JSON and contains source, license, provenance, tier, spot_checks, spot_check_summary"
  ]
}