{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "arrayShape": "cr:arrayShape",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "containedIn": "cr:containedIn",
    "cr": "http://mlcommons.org/croissant/",
    "data": {
      "@id": "cr:data",
      "@type": "@json"
    },
    "dataType": {
      "@id": "cr:dataType",
      "@type": "@vocab"
    },
    "dct": "http://purl.org/dc/terms/",
    "equivalentProperty": "cr:equivalentProperty",
    "examples": "cr:examples",
    "extract": "cr:extract",
    "field": "cr:field",
    "fileObject": "cr:fileObject",
    "fileProperty": "cr:fileProperty",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isArray": "cr:isArray",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "parentField": "cr:parentField",
    "path": "cr:path",
    "prov": "http://www.w3.org/ns/prov#",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "replace": "cr:replace",
    "samplingRate": "cr:samplingRate",
    "sc": "https://schema.org/",
    "separator": "cr:separator",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform"
  },
  "@type": "sc:Dataset",
  "conformsTo": [
    "http://mlcommons.org/croissant/1.1",
    "http://mlcommons.org/croissant/RAI/1.0"
  ],
  "name": "OpenWhistle-Detection-Finetuning",
  "alternateName": [
    "dolphinteam/OpenWhistle-Detection-Finetuning"
  ],
  "description": "OpenWhistle Detection Finetuning is a public whistle-type detection finetuning dataset. It contains 5,600 fixed-length audio windows with audio, a seven-dimensional multi-label whistle-type target vector, a derived binary whistle/noise label, original source labels, and original clip filenames. Background/no-whistle windows are represented by an all-zero target vector. The train/validation/test split is session-disjoint, uses ratios 0.70/0.15/0.15, and is exactly balanced across whistle-type source labels and background examples within each split.",
  "url": "https://huggingface.co/datasets/dolphinteam/OpenWhistle-Detection-Finetuning",
  "citeAs": "Faadil Mustun, Chiara Semenzin, Roberto Dessi, Pablo Robin Guerrero, Pierre Orhan, Alexis Emanuelli, Emanuele Rossi, Yair Lakretz, Gonzalo de Polavieja, and German Sumbre. OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations. NeurIPS 2026 Datasets & Evaluations Track (Spotlight), arXiv:2609.34839. https://arxiv.org/abs/2609.34839",
  "creator": [
    {
      "@type": "Person",
      "name": "Faadil Mustun"
    },
    {
      "@type": "Person",
      "name": "Chiara Semenzin"
    },
    {
      "@type": "Person",
      "name": "Roberto Dessi"
    },
    {
      "@type": "Person",
      "name": "Pablo Robin Guerrero"
    },
    {
      "@type": "Person",
      "name": "Pierre Orhan"
    },
    {
      "@type": "Person",
      "name": "Alexis Emanuelli"
    },
    {
      "@type": "Person",
      "name": "Emanuele Rossi"
    },
    {
      "@type": "Person",
      "name": "Yair Lakretz"
    },
    {
      "@type": "Person",
      "name": "Gonzalo de Polavieja"
    },
    {
      "@type": "Person",
      "name": "German Sumbre"
    }
  ],
  "publisher": {
    "@type": "Organization",
    "name": "Dolphin Team",
    "url": "https://huggingface.co/dolphinteam"
  },
  "datePublished": "2026-04-22",
  "dateModified": "2026-09-29",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "identifier": "https://huggingface.co/datasets/dolphinteam/OpenWhistle-Detection-Finetuning",
  "sameAs": [
    "https://arxiv.org/abs/2609.34839",
    "https://huggingface.co/collections/dolphinteam/neurips26-openwhistle"
  ],
  "keywords": [
    "dolphin",
    "bioacoustics",
    "whistle-detection",
    "audio",
    "multi-label detection",
    "passive acoustic monitoring",
    "session-disjoint split",
    "Hugging Face",
    "Croissant",
    "Responsible AI"
  ],
  "isAccessibleForFree": true,
  "isLiveDataset": false,
  "spatialCoverage": {
    "@type": "Place",
    "name": "Dolphin Reef, Eilat, northern Gulf of Aqaba"
  },
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "repo",
      "name": "repo",
      "description": "The Hugging Face git repository containing the dataset parquet conversion.",
      "contentUrl": "https://huggingface.co/datasets/dolphinteam/OpenWhistle-Detection-Finetuning/tree/refs%2Fconvert%2Fparquet",
      "encodingFormat": "git+https",
      "sha256": "https://github.com/mlcommons/croissant/issues/80"
    },
    {
      "@type": "cr:FileSet",
      "@id": "parquet-files-for-config-default",
      "containedIn": {
        "@id": "repo"
      },
      "encodingFormat": "application/x-parquet",
      "includes": "default/*/*.parquet"
    }
  ],
  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "default_splits",
      "name": "default_splits",
      "description": "Splits for the default config.",
      "dataType": "cr:Split",
      "key": {
        "@id": "default_splits/split_name"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "default_splits/split_name",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "default_splits/split_name": "train"
        },
        {
          "default_splits/split_name": "validation"
        },
        {
          "default_splits/split_name": "test"
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "default",
      "name": "default",
      "description": "Session-disjoint multi-label whistle-type detection benchmark with train, validation, and test splits.",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "default/split",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "fileProperty": "fullpath"
            },
            "transform": {
              "regex": "default/(?:partial-)?(train|validation|test)/.+parquet$"
            }
          },
          "references": {
            "field": {
              "@id": "default_splits/split_name"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "default/audio",
          "dataType": "sc:AudioObject",
          "samplingRate": 96000,
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "column": "audio"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "default/binary_label",
          "dataType": "sc:Integer",
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "column": "binary_label"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "default/name",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "column": "name"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "default/source_label",
          "dataType": "sc:Text",
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "column": "source_label"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "default/label",
          "dataType": "cr:Int64",
          "isArray": true,
          "arrayShape": "7",
          "source": {
            "fileSet": {
              "@id": "parquet-files-for-config-default"
            },
            "extract": {
              "column": "label"
            }
          }
        }
      ]
    }
  ],
  "rai:dataCollection": "The dataset was assembled from curated OpenWhistle detection windows derived from passive acoustic monitoring at Dolphin Reef, Eilat, a semi-natural marine site on the northern Gulf of Aqaba. Fixed hydrophones recorded a stable pod of known dolphins in a large natural marine area open to the sea. The released examples are fixed-length audio windows labeled for whistle-type detection, with a derived binary whistle/noise label retained for convenience.",
  "rai:dataCollectionType": [
    "Physical data collection",
    "Direct measurement",
    "Manual Human Curator",
    "Software Collection"
  ],
  "rai:dataCollectionRawData": "Raw data consisted of underwater acoustic recordings captured by fixed and hidden hydrophones. The released detection dataset contains curated windows and metadata fields: audio, binary_label, name, source_label, and label. The label field is a seven-dimensional multi-label tensor ordered as SW_Neo, SW_Luna, SW_Nikita, SW_Nana, SW_Yosefa, SW_Dana, NSW_1; background/no-whistle examples have an all-zero tensor.",
  "rai:dataCollectionMissingData": "The public dataset does not include per-example video context, exact hydrophone coordinates, environmental conditions, full behavioral context, or human-identifying information. The task labels identify which retained whistle types are present in each window, with all-zero targets for background; the dataset does not provide full temporal localization inside longer recordings, behavioral meaning, or exhaustive annotation of all possible dolphin vocalization types.",
  "rai:dataPreprocessingProtocol": [
    "Curated OpenWhistle detection windows were encoded as a multi-label whistle-type benchmark: the label tensor records the whistle types present in the order SW_Neo, SW_Luna, SW_Nikita, SW_Nana, SW_Yosefa, SW_Dana, NSW_1, while background windows use an all-zero vector.",
    "The original source labels were retained in source_label, and binary_label was derived as whistle when at least one label dimension is active and noise otherwise.",
    "Train, validation, and test splits were assigned by whole recording sessions using exact MILP source-label balancing with seed 42. This produced 3,920 train rows, 840 validation rows, and 840 test rows.",
    "The final split has 400 examples per whistle type across seven classes and 2,800 background examples overall, with exact per-split targets and zero pairwise session overlap across all splits."
  ],
  "rai:dataAnnotationProtocol": "The dataset uses curated detection windows with source labels indicating background/noise or one of seven whistle categories. The primary target is the seven-dimensional whistle-type label tensor; binary_label is derived from this tensor for convenience. The source_label field is retained to preserve the original source category.",
  "rai:annotationsPerItem": "Each item has a seven-dimensional whistle-type label tensor. Background/no-whistle items have an all-zero tensor. Items also include a derived binary whistle/noise label and the original source_label.",
  "rai:machineAnnotationTools": [
    "Exact MILP assignment was used to create session-disjoint, source-label-balanced train/validation/test splits.",
    "Hugging Face Datasets and parquet conversion tooling were used to package and publish the dataset."
  ],
  "rai:dataUseCases": "[\"Construct represented: curated short underwater acoustic windows labeled with the dolphin whistle types present, with all-zero targets for background/no-whistle windows.\", \"Validated use case: training, validation, and testing of multi-label dolphin whistle-type detection models on curated OpenWhistle windows.\", \"Validated use case: evaluating detection models under a session-disjoint protocol with exact balance of whistle-type source labels and background examples across train, validation, and test splits.\", \"Validated use case: benchmarking audio representation models for whistle-type detection before downstream whistle-type classification.\", \"Supported audit use case: inspecting source-label coverage using the retained source_label and label tensor fields.\", \"Not validated use case: full continuous-recording segmentation, behavioral interpretation, individual identification, ecological decision-making, or deployment performance in other sites without additional local validation.\"]",
  "rai:dataLimitations": "[\"The dataset is designed for multi-label whistle-type detection on fixed-length curated windows, not for full continuous-recording segmentation, behavioral interpretation, individual identification, or ecological decision-making.\", \"The clips come from a semi-natural site with a stable pod of known dolphins, so performance may not generalize to fully wild populations, other species, other recording devices, or different acoustic environments without validation.\", \"Exact balancing of whistle-type source labels and background windows improves benchmark comparability but does not represent natural class frequencies in continuous recordings.\", \"The source_label tensor tracks seven whistle categories only and should not be interpreted as exhaustive coverage of all possible dolphin vocalization types.\", \"Session-disjoint splitting avoids session leakage across train/validation/test, but users should preserve the provided splits when reporting benchmark results.\"]",
  "rai:dataBiases": "[\"The dataset reflects Dolphin Reef recordings and may overrepresent acoustic and social conditions at that site.\", \"The dataset is intentionally balanced: background windows and whistle-type examples are equally represented overall, and each retained source whistle label has exact split targets. This reduces evaluation imbalance but differs from natural recording distributions.\", \"Curated clip selection may overrepresent clear, segmentable, or higher-quality examples and underrepresent ambiguous, overlapping, low-SNR, or rare acoustic events.\", \"Aggregate detection metrics can hide differences between whistle categories, so performance should be inspected by source label as well as overall.\"]",
  "rai:personalSensitiveInformation": "[\"No human subjects are involved in the dataset. Public columns contain animal audio clips, binary/source labels, original clip filenames, and source-label tensors.\", \"The data consists of underwater animal acoustic recordings. Users should nevertheless treat incidental non-target sounds as possible and avoid attempts to infer people, vessels, precise locations, or sensitive field-site details from the audio or filenames.\"]",
  "rai:dataSocialImpact": "The dataset supports open research on passive acoustic monitoring, dolphin bioacoustics, and non-invasive whistle-type detection. Positive impacts include reproducible evaluation of whistle-type detectors and improved tooling for animal communication research. Misuse risks include deploying detectors in new sites without local validation, treating benchmark detection as proof of ecological or behavioral conclusions, or ignoring source-label-specific failure modes hidden by aggregate metrics. The original data collection was passive and non-invasive, did not interfere with dolphin behavior, and involved no human subjects.",
  "rai:hasSyntheticData": false,
  "rai:dataReleaseMaintenancePlan": "The dataset is hosted publicly on Hugging Face under dolphinteam/OpenWhistle-Detection-Finetuning and released under CC-BY 4.0. The release documents the exact session-disjoint split, source-label balancing, seed, and leakage checks. Future OpenWhistle releases may expand detection coverage, labels, contextual metadata, or evaluation protocols.",
  "prov:wasDerivedFrom": [
    {
      "@id": "urn:openwhistle:detection-source-recordings-and-clips:neurips-2026",
      "prov:label": "OpenWhistle curated detection clips and passive acoustic source recordings",
      "sc:license": "https://creativecommons.org/licenses/by/4.0/",
      "prov:wasAttributedTo": {
        "@id": "openwhistle_authors",
        "prov:label": "OpenWhistle authors"
      }
    }
  ],
  "prov:wasGeneratedBy": [
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q4929239"
      },
      "prov:label": "Passive acoustic data collection",
      "sc:description": "Underwater audio was collected passively using fixed and hidden hydrophones. The collection did not interfere with dolphin behavior, did not train or constrain animals, and involved no human subjects.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:Agent",
          "@id": "openwhistle_authors",
          "prov:label": "OpenWhistle authors",
          "sc:description": "OpenWhistle research team responsible for the OpenWhistle data release and documentation."
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q109719325"
      },
      "prov:label": "Detection window labeling and target encoding",
      "sc:description": "Curated detection windows were assigned source labels and encoded as seven-dimensional whistle-type target vectors. Background/no-whistle windows use an all-zero vector, and binary_label is derived from whether any target dimension is active.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:Agent",
          "@id": "openwhistle_authors",
          "prov:label": "OpenWhistle authors",
          "sc:description": "OpenWhistle research team responsible for curation and whistle-type target encoding."
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q5227332"
      },
      "prov:label": "Session-disjoint split optimization and dataset packaging",
      "sc:description": "Whole recording sessions were assigned to train, validation, and test splits using exact MILP source-label balancing with seed 42. The final dataset contains 5,600 rows, 241 sessions, zero pairwise session overlap across splits, and exact whistle-type/background targets. The data was packaged as Hugging Face parquet files.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "exact_milp_source_label_balancing",
          "prov:label": "Exact MILP source-label balancing",
          "sc:description": "Optimization procedure used to assign whole sessions while matching split ratios and label targets."
        },
        {
          "@type": "prov:SoftwareAgent",
          "@id": "hugging_face_datasets",
          "prov:label": "Hugging Face Datasets",
          "sc:description": "Software tooling used to package and publish the dataset with Audio and sequence features."
        },
        {
          "@type": "prov:Agent",
          "@id": "openwhistle_authors",
          "prov:label": "OpenWhistle authors",
          "sc:description": "OpenWhistle research team responsible for preprocessing, curation, and release."
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q3306762"
      },
      "prov:label": "Leakage and split audit",
      "sc:description": "The released artifacts document pairwise session overlap checks. The final split has train__validation=0, train__test=0, and validation__test=0 session overlap.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:Agent",
          "@id": "openwhistle_authors",
          "prov:label": "OpenWhistle authors",
          "sc:description": "OpenWhistle research team responsible for split audit and release artifacts."
        }
      ]
    }
  ]
}