{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "data": {
      "@id": "cr:data",
      "@type": "@json"
    },
    "dataType": {
      "@id": "cr:dataType",
      "@type": "@vocab"
    },
    "dct": "http://purl.org/dc/terms/",
    "examples": {
      "@id": "cr:examples",
      "@type": "@json"
    },
    "equivalentProperty": "cr:equivalentProperty",
    "samplingRate": "cr:samplingRate",
    "extract": "cr:extract",
    "field": "cr:field",
    "fileProperty": "cr:fileProperty",
    "fileObject": "cr:fileObject",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "parentField": "cr:parentField",
    "path": "cr:path",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "replace": "cr:replace",
    "sc": "https://schema.org/",
    "separator": "cr:separator",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform"
  },
  "@type": "sc:Dataset",
  "conformsTo": [
    "http://mlcommons.org/croissant/1.0",
    "http://mlcommons.org/croissant/RAI/1.0"
  ],
  "name": "SynBench",
  "description": "Differentially private synthetic text from six domain-specific classification datasets (biomedical, finance, legal), generated by DP-Gen, AUG-PE, DP-Dfs and EPSVec at epsilon in {0.5, 1, 2, 4, inf}, with the original splits and membership-inference audit data.",
  "url": "https://huggingface.co/datasets/Ydsun/SynBench",
  "version": "1.0.0",
  "datePublished": "2026",
  "license": [
    "https://creativecommons.org/licenses/by-nc-sa/4.0/",
    "https://creativecommons.org/licenses/by-sa/4.0/",
    "https://opensource.org/licenses/MIT",
    "https://www.gnu.org/licenses/gpl-3.0.html"
  ],
  "creator": [
    {
      "@type": "sc:Person",
      "name": "Yidan Sun"
    },
    {
      "@type": "sc:Person",
      "name": "Viktor Schlegel"
    },
    {
      "@type": "sc:Person",
      "name": "Srinivasan Nandakumar"
    },
    {
      "@type": "sc:Person",
      "name": "Iqra Zahid"
    },
    {
      "@type": "sc:Person",
      "name": "Yuping Wu"
    },
    {
      "@type": "sc:Person",
      "name": "Yulong Wu"
    },
    {
      "@type": "sc:Person",
      "name": "Hao Li"
    },
    {
      "@type": "sc:Person",
      "name": "Jie Zhang"
    },
    {
      "@type": "sc:Person",
      "name": "Warren Del-Pinto"
    },
    {
      "@type": "sc:Person",
      "name": "Goran Nenadic"
    },
    {
      "@type": "sc:Person",
      "name": "Siew Kei Lam"
    },
    {
      "@type": "sc:Person",
      "name": "Anil Anthony Bharath"
    }
  ],
  "citeAs": "@inproceedings{sun2026synbench, title={SynBench: A Benchmark for Differentially Private Text Generation}, author={Sun, Yidan and Schlegel, Viktor and others}, booktitle={NeurIPS Evaluations and Datasets Track}, year={2026}}",
  "keywords": [
    "differential privacy",
    "synthetic data",
    "text generation",
    "membership inference",
    "benchmark"
  ],
  "isAccessibleForFree": true,
  "rai:dataCollection": "No new data was collected. Original records come from public datasets (hallmarks_of_cancer: bigbio/hallmarks_of_cancer (gpl-3.0); DMSAFN: Daniel-ML/sentiment-analysis-for-financial-news-v2 (cc-by-nc-sa-4.0); Lc37Fnd: luckycat37/financial-news-dataset (cc-by-nc-sa-4.0); ZSTFNS: zeroshot/twitter-financial-news-sentiment (mit); AYSLAX: clairebarale/AsyLex (cc-by-nc-sa-4.0); EURLEX: EURLEX57K (cc-by-sa-4.0)). Synthetic records were generated from each train split by four DP text generators; MIA data was sampled from the train split as described in the paper.",
  "rai:dataCollectionType": [
    "Secondary Data Analysis",
    "Software Collection"
  ],
  "rai:dataPreprocessingProtocol": [
    "OLMo-2 DP-Gen outputs de-duplicated on (text, label), keeping the first occurrence.",
    "EURLEX restricted to 8 EuroVoc concepts of EURLEX57K; LC37FND labels derived from source sentiment scores.",
    "Original texts are otherwise unchanged from the sources."
  ],
  "rai:dataUseCases": "Benchmarking DP synthetic text generation for downstream utility, distributional fidelity and empirical privacy leakage.",
  "rai:dataLimitations": "English only; six datasets; generators of 1-8B parameters. Several originals likely appear in LLM pre-training corpora, which can inflate utility and leakage. Some low-epsilon DP-Dfs outputs are degenerate.",
  "rai:dataBiases": "Inherits the topical, temporal and labelling biases of the source datasets (financial news and tweets, EU legislation, Canadian refugee decisions, biomedical abstracts).",
  "rai:personalSensitiveInformation": "ASYLAX contains publicly released, anonymised refugee-case decisions; other sources are news, tweets, legislation and scientific abstracts. Datasets with patient data (PsyTAR, N2C2'08, MIMIC) are excluded.",
  "rai:dataSocialImpact": "Standardised evaluation of DP text generation can support safer sharing of sensitive text; the MIA audit could also inform attackers.",
  "rai:hasSyntheticData": true,
  "rai:dataReleaseMaintenancePlan": "Maintained by the authors on Hugging Face; versions are tagged.",
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "repo",
      "name": "repo",
      "description": "The Hugging Face git repository.",
      "contentUrl": "https://huggingface.co/datasets/Ydsun/SynBench",
      "encodingFormat": "git+https",
      "sha256": "main"
    },
    {
      "@type": "cr:FileSet",
      "@id": "original-files",
      "name": "original-files",
      "description": "Real train/test splits of each dataset.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/original/*-original.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "synthetic-files",
      "name": "synthetic-files",
      "description": "Synthetic datasets, one file per (dataset, generator, method, epsilon).",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/*/*/train*.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "mia-private-files",
      "name": "mia-private-files",
      "description": "100 private subsets per dataset sampled for the MIA audit.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/mia/D_sample/sampled_private_datasets/dataset_*.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "mia-reference-files",
      "name": "mia-reference-files",
      "description": "4 auxiliary reference sets per dataset; each target is in exactly two.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/mia/D_sample/sampled_4ref_datasets_outliers/dataset_*.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "mia-synthetic-files",
      "name": "mia-synthetic-files",
      "description": "Synthetic version of every private subset, per method and epsilon.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/mia/D_synth/*/finalresults/dataset_*_samples.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "mia-target-files",
      "name": "mia-target-files",
      "description": "Outlier records used as MIA targets.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/jsonlines",
      "includes": "*/mia/vulsamples/all_outliers.jsonl"
    },
    {
      "@type": "cr:FileSet",
      "@id": "mia-membership-labels",
      "name": "mia-membership-labels",
      "description": "sampling_labels.json: for each sampled set, whether and which target was inserted.",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/json",
      "includes": "*/mia/D_sample/*/sampling_labels.json"
    }
  ],
  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "original",
      "name": "original",
      "description": "Real records. The test split is held out and never used for generation.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "original/dataset",
          "name": "dataset",
          "description": "Source dataset (config name).",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "original-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/original/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "original/split",
          "name": "split",
          "description": "train or test.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "original-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/(train|test)-original\\.jsonl$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "original/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "original-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "original/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "original-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "synthetic",
      "name": "synthetic",
      "description": "DP synthetic records generated from the train split.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "synthetic/dataset",
          "name": "dataset",
          "description": "Source dataset (config name).",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/[^/]+/[^/]+/train[^/]*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "synthetic/generator",
          "name": "generator",
          "description": "Generator folder, e.g. Llama-3.2-1B.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?[^/]+/([^/]+)/[^/]+/train[^/]*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "synthetic/method",
          "name": "method",
          "description": "dp-transformers (DP-Gen), aug-pe, dp-diffusion or EPSVec.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?[^/]+/[^/]+/([^/]+)/train[^/]*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "synthetic/epsilon",
          "name": "epsilon",
          "description": "Privacy budget from the file name: 0.5, 1, 2, 4; 'nodp' or '0' = no DP.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/train(0\\.5|1|2|4|nodp|0)-[^/]*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "synthetic/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "synthetic/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "synthetic-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "mia_private",
      "name": "mia_private",
      "description": "Private subsets D_prv^i; each includes a random outlier target with probability 0.5.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "mia_private/dataset",
          "name": "dataset",
          "description": "Source dataset.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-private-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/mia/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_private/subset_id",
          "name": "subset_id",
          "description": "Index i of the subset.",
          "dataType": "sc:Integer",
          "source": {
            "fileSet": {
              "@id": "mia-private-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/dataset_([0-9]+)\\.jsonl$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_private/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-private-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_private/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-private-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "mia_reference",
      "name": "mia_reference",
      "description": "Reference sets D_ref^1..4 built from the auxiliary split.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "mia_reference/dataset",
          "name": "dataset",
          "description": "Source dataset.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-reference-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/mia/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_reference/ref_id",
          "name": "ref_id",
          "description": "Index of the reference set.",
          "dataType": "sc:Integer",
          "source": {
            "fileSet": {
              "@id": "mia-reference-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/dataset_([0-9]+)\\.jsonl$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_reference/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-reference-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_reference/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-reference-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "mia_synthetic",
      "name": "mia_synthetic",
      "description": "Synthetic version of each private subset.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/dataset",
          "name": "dataset",
          "description": "Source dataset.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/mia/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/method",
          "name": "method",
          "description": "augpe or dp-transformers.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/D_synth/([^/]+)/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/subset_id",
          "name": "subset_id",
          "description": "Index i of the private subset.",
          "dataType": "sc:Integer",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*/dataset_([0-9]+)_e[^/]*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/epsilon",
          "name": "epsilon",
          "description": "Privacy budget; 0 = no DP.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^.*_e([0-9.]+)_samples\\.jsonl$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_synthetic/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-synthetic-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "mia_targets",
      "name": "mia_targets",
      "description": "Outlier records (top-1% distance and LOF outliers) used as MIA targets.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "mia_targets/dataset",
          "name": "dataset",
          "description": "Source dataset.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-target-files"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "^(?:.*/)?([^/]+)/mia/.*$"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_targets/index",
          "name": "index",
          "description": "Row index in original/train-original.jsonl.",
          "dataType": "sc:Integer",
          "source": {
            "fileSet": {
              "@id": "mia-target-files"
            },
            "extract": {
              "column": "index"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_targets/text",
          "name": "text",
          "description": "Record text.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-target-files"
            },
            "extract": {
              "column": "text"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "mia_targets/label",
          "name": "label",
          "description": "Label(s) of the record; single-label datasets use one element.",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "mia-target-files"
            },
            "extract": {
              "column": "label"
            }
          },
          "repeated": true
        }
      ]
    }
  ]
}