dolphinteam/NewPretraining
收藏资源简介:
--- dataset_info: features: - name: audio dtype: audio: sampling_rate: 44100 - name: start_time dtype: float32 - name: end_time dtype: float32 - name: duration dtype: float32 - name: year dtype: int32 - name: hydrophone dtype: string splits: - name: train num_bytes: 64727228124.0 num_examples: 28410 - name: validation num_bytes: 7695460500.0 num_examples: 3370 download_size: 33930289987 dataset_size: 72422688624.0 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* --- # NewPretraining Cleaned version of `dolphinteam/Dolphin-Pretraining-clean-year-hydro` with the segments overlapping the whistle finetuning dataset removed. ## Notes - Audio is embedded directly inside the Parquet shards. - Shards are written with `SNAPPY` compression for the audio columns. - The split strategy is deterministic at the recording level with seed `42`.




