{"id":902650,"date":"2026-05-02T20:59:05","date_gmt":"2026-05-02T20:59:05","guid":{"rendered":"https:\/\/example.com\/?p=1002"},"modified":"2026-06-10T02:34:02","modified_gmt":"2026-06-10T02:34:02","slug":"inside-the-data-2609-files","status":"publish","type":"post","link":"https:\/\/cri-one.com\/blog\/2026\/05\/02\/inside-the-data-2609-files\/","title":{"rendered":"Inside the Data: 2,609 Files, Millions of Compounds"},"content":{"rendered":"

What “the data” actually is<\/h2>\n

It’s easy to throw around big numbers \u2014 millions of compounds, thousands of procedures \u2014 without explaining what they refer to. This post unpacks the dataset so you can decide whether it fits your work.<\/p>\n

The shape of the database<\/h2>\n

The core probability database lives in a directory called bluray_data<\/code>. It contains 2,609 data files<\/strong> spanning four broad categories:<\/p>\n

    \n
  1. Compound combinations<\/strong> \u2014 millions of element pairings and groupings, indexed for fast lookup. You can search by element symbols (for example, ['H', 'O']<\/code>) and get ranked candidate compounds back.<\/li>\n
  2. Probability matrices<\/strong> \u2014 statistical data on how likely each combination is to form, behave a certain way, or persist. This is the layer that lets you rank rather than just retrieve.<\/li>\n
  3. Transformation database<\/strong> \u2014 a “how to” library of structured procedures for converting one compound into another. Thousands of synthesis methods, organized so that a query returns a usable path rather than a paper to read.<\/li>\n
  4. Discovery reports<\/strong> \u2014 detailed compound analyses, the equivalent of having a reference write-up already done for many of the entries you’ll touch.<\/li>\n<\/ol>\n

    Formats and access<\/h2>\n

    The dataset ships in multiple formats so it can drop into whatever stack you’re already using:<\/p>\n

      \n
    • CSV<\/strong> for spreadsheet and pandas workflows.<\/li>\n
    • JSON<\/strong> for programmatic access and web tooling.<\/li>\n
    • Compressed formats<\/strong> for moving the full corpus efficiently.<\/li>\n<\/ul>\n

      Alongside the raw files, there is a processed-output directory (alchemy_data_v2_output<\/code>) containing pre-computed views, plus a core_systems<\/code> directory with the components that wire everything together.<\/p>\n

      Documentation, not just data<\/h2>\n

      One of the things that distinguishes this package from “a folder of CSVs” is the documentation layer \u2014 61 files covering quick starts, architecture notes, generation guides, and validation results:<\/p>\n

        \n
      • QUICK_START_USING_DATA.md<\/code> \u2014 the fast on-ramp.<\/li>\n
      • USE_BLURAY_DATA.md<\/code> \u2014 full usage reference.<\/li>\n
      • HOW_TO_RUN_BLURAY_GENERATOR.md<\/code> \u2014 for teams that want to extend the dataset.<\/li>\n
      • METHOD_DATA_SYSTEM_README.md<\/code> \u2014 system architecture.<\/li>\n
      • TEST_ANALYSIS.md<\/code> \u2014 validation results.<\/li>\n<\/ul>\n

        Why this organization matters<\/h2>\n

        Datasets become useful when the structure makes the next question cheap to ask. Probability matrices sit next to the combinations they rank. Transformation guides are linked to the compounds they produce. Discovery reports are reachable from the compound that prompted them. The intent is that a researcher can move from a question to an answer without rebuilding the index every time.<\/p>\n

        The next post turns from “what is in here” to “what does it do for you” \u2014 a tour of the use cases that drove the design.<\/p>\n

        <\/p>\n

        \n