Update croissant_rai_OpenWhistle-Classification-Finetuning.json

croissant_rai_OpenWhistle-Classification-Finetuning.json

@@ -71,7 +71,7 @@
     "url": "https://huggingface.co/OpenWhistleNeurIPS26"
   },
   "datePublished": "2026-04-20",
-  "dateModified": "2026-05-05",
+  "dateModified": "2026-05-06",
   "license": "https://creativecommons.org/licenses/by/4.0/",
   "version": "1.0.0",
   "identifier": "https://huggingface.co/datasets/OpenWhistleNeurIPS26/OpenWhistle-Classification-Finetuning",
@@ -1769,10 +1769,10 @@
     "Manual Human Curator",
     "Software Collection"
   ],
-  "rai:dataCollectionRawData": "Raw data consisted of underwater acoustic recordings captured by fixed and hidden hydrophones. The released classification dataset contains short whistle clips and metadata fields: audio, label, name, onset, offset, duration, recording_duration, whistle_type, whistle_name, f0_time, f0_hz, f0_conf, f0_ok, f0_bad_reason, and f0_spectrogram.",
+  "rai:dataCollectionRawData": "Raw data consisted of underwater acoustic recordings captured by fixed and hidden hydrophones. The released classification dataset contains short whistle clips and metadata fields: audio, label, name, onset, offset, duration, recording_duration, whistle_type, whistle_name, f0_time, f0_hz, f0_conf, f0_ok, f0_bad_reason, and f0_spectrogram. The all and all-review-sample configurations additionally include snr_db, an estimated clip-level signal-to-noise ratio in dB.",
   "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 balanced configuration intentionally excludes rare whistle categories to support reliable six-class evaluation, while all and unbalanced configurations retain a broader ten-class label space with imbalanced classes.",
   "rai:dataPreprocessingProtocol": [
-    "Whistle clips were extracted from OpenWhistle recordings and packaged as Hugging Face parquet datasets with audio, timing metadata, F0 contours, F0 confidence values, and rendered F0 spectrogram images.",
+    "Whistle clips were extracted from OpenWhistle recordings and packaged as Hugging Face parquet datasets with audio, timing metadata, F0 contours, F0 confidence values, rendered F0 spectrogram images, and, for the all and all-review-sample configurations, clip-level snr_db estimates.",
     "The main balanced configuration uses six classes and exact class balancing with 2,100 train, 450 validation, and 450 test examples. Splits are session-disjoint, so no recording session appears in more than one split.",
     "The unbalanced and all configurations expose ten whistle classes for additional analysis. Review-sample configurations were generated deterministically after split creation and are intended for manual inspection, not reporting."
   ],
@@ -1780,7 +1780,7 @@
   "rai:annotationsPerItem": "Each released example has one whistle-type class label. The label corresponds to either a signature whistle class or a non-signature whistle class, depending on the configuration.",
   "rai:machineAnnotationTools": [
     "ARTwarp and dynamic time warping were used in the annotation pipeline to cluster whistle contours before expert refinement.",
-    "F0 extraction and spectrogram rendering software were used to create f0_time, f0_hz, f0_conf, f0_ok, f0_bad_reason, and f0_spectrogram fields.",
+    "F0 extraction and spectrogram rendering software were used to create f0_time, f0_hz, f0_conf, f0_ok, f0_bad_reason, and f0_spectrogram fields. Signal-to-noise estimation software was used to create snr_db for the all and all-review-sample configurations.",
     "Hugging Face Datasets tooling was used to package and publish the dataset as parquet configurations."
   ],
   "rai:dataUseCases": "[\"Construct represented: short dolphin whistle clips with expert-refined whistle-type labels, timing metadata, F0 tracks, and F0 spectrograms.\", \"Validated use case: training, validation, and testing of dolphin whistle-type classifiers on the balanced six-class OpenWhistle benchmark with session-disjoint train, validation, and test splits.\", \"Validated use case: linear probing or finetuning of audio representation models for fine-grained intra-species dolphin whistle discrimination.\", \"Supported exploratory use case: analysis of ten-class whistle categories using the all and unbalanced configurations, with explicit attention to class imbalance.\", \"Supported reviewer use case: manual inspection through deterministic review-sample configurations.\", \"Not validated use case: inferring behavioral intent, communicative meaning, animal welfare state, individual identity beyond the released label definitions, or deployment performance in other sites without additional biological validation.\"]",
@@ -1846,7 +1846,7 @@
         "@id": "https://www.wikidata.org/wiki/Q5227332"
       },
       "prov:label": "Feature extraction and dataset packaging",
-      "sc:description": "Whistle clips, class labels, timing metadata, F0 contours, F0 confidence values, and rendered F0 spectrograms were packaged into Hugging Face parquet configurations. The balanced, unbalanced, and all configurations were split into train, validation, and test sets with session-disjoint assignment.",
+      "sc:description": "Whistle clips, class labels, timing metadata, F0 contours, F0 confidence values, rendered F0 spectrograms, and clip-level snr_db estimates where available were packaged into Hugging Face parquet configurations. The balanced, unbalanced, and all configurations were split into train, validation, and test sets with session-disjoint assignment.",
       "prov:wasAttributedTo": [
         {
           "@type": "prov:SoftwareAgent",
@@ -1879,4 +1879,4 @@
       ]
     }
   ]
-}
+}