{
  "benchmark_final_hf_columns": [
    {
      "name": "Unique_identifier",
      "type": "string",
      "description": "Unique row identifier, e.g. USPTO_US03930836_0. Use this column to join split files."
    },
    {
      "name": "OriginalReaction",
      "type": "string",
      "description": "Original USPTO reaction SMILES before any curation or canonicalization."
    },
    {
      "name": "InputReaction",
      "type": "string",
      "description": "Model input: atom-imbalanced reaction SMILES with reagents moved to the reactant side (reactants.reagents>>incomplete_product)."
    },
    {
      "name": "TargetReaction",
      "type": "string",
      "description": "Ground-truth target: full atom- and charge-balanced reaction SMILES from FlowER."
    },
    {
      "name": "TargetReaction_with_stereo",
      "type": "string",
      "description": "Target reaction with stereochemistry transferred back from the original USPTO record. Null where stereo could not be resolved."
    },
    {
      "name": "has_unresolved_stereo",
      "type": "boolean",
      "description": "True if one or more stereocentres in the FlowER reaction could not be resolved from the USPTO record."
    },
    {
      "name": "CanonicalOriginalReaction",
      "type": "string",
      "description": "Canonicalized original USPTO reaction in standard reactants>reagents>products format, no atom mapping."
    },
    {
      "name": "CanonicalReaction",
      "type": "string",
      "description": "Canonicalised reaction with reagents identified based on atom mapping (similar to USPTO Stereo) "
    },
    {
      "name": "MissingNrAtoms",
      "type": "integer",
      "description": "Number of heavy atoms present in the target but missing from the input product side. Primary difficulty axis. Based on CanonicalReaction to avoid counting reagents as missing atoms."
    },
    {
      "name": "MissingNrCarbons",
      "type": "integer",
      "description": "Number of carbon atoms present in the target but missing from the input product side.  Based on CanonicalReaction to avoid counting reagents."
    },
    {
      "name": "MissingMw",
      "type": "float",
      "description": "Molecular weight (Da) of the missing fragment (target minus input product). Based on CanonicalReaction to avoid counting reagents."
    },
    {
      "name": "IsBalanced",
      "type": "boolean",
      "description": "True if the CanonicalReaction record is already atom-balanced before curation."
    },
    {
        "name": "Balance",
        "type": "string",
        "description": "String that describes the type and number of missing atoms."
    },
    {
      "name": "rxn_class",
      "type": "string",
      "description": "USPTO reaction class label (NameRxn classification)."
    },
    {
      "name": "1kTPLLabel",
      "type": "string",
      "description": "Reaction template label from the 1 000 most common templates. Null if the reaction does not match any of the top-1k templates."
    },
    {
      "name": "FlowER_ID",
      "type": "string",
      "description": "Identifier of the linked FlowER mechanistic reaction step. Used to trace provenance back to the FlowER dataset."
    },
    {
      "name": "PatentNumber",
      "type": "string",
      "description": "USPTO patent number from which the reaction was extracted."
    },
    {
      "name": "ParagraphNum",
      "type": "integer",
      "description": "Paragraph number within the patent from which the reaction was extracted."
    },
    {
      "name": "Year",
      "type": "integer",
      "description": "Year the USPTO patent was filed."
    }
  ],
  "split_files": {
    "description": "Each split file contains Unique_identifier, row_index, and a split assignment column. Join on Unique_identifier. Values: 0=train, 1=valid, 2=test.",
    "files": [
      {
        "folder": "random_splits/",
        "split_column": "random_split",
        "description": "5 random train/valid/test splits (rep0-rep4). Seeds 42-46."
      },
      {
        "folder": "group_splits/",
        "split_column": "group_split",
        "description": "5 group-OOD splits based on DRFP fingerprint clustering. Test set contains reaction groups not seen during training."
      },
      {
        "folder": "extreme_ood_splits/",
        "split_column": "extreme_ood_split",
        "description": "5 extreme OOD splits using KNN-graph separation. Test set is maximally dissimilar from the training set by Tanimoto distance."
      }
    ]
  }
}