AE-W/generative-sound-masking-retrieval-dasheng-v1
Generative Sound Masking — DaSheng retrieval v1 This dataset stores exhaustive top-15 retrieval from the 237,500 prompt–seed candidate audios for 48,840 background audios. One row per background contains 15 ranked candidates, including prompt, seed, clip ID, audio SHA256 and cosine score. Source audio stays in the two original public datasets; no audio is duplicated here. This is an incremental run. Read progress.json before treating it as complete. The data/train split is an… See the full description on the dataset page: https://huggingface.co/datasets/AE-W/generative-sound-masking-retrieval-dasheng-v1.
Generative Sound Masking — DaSheng retrieval v1
This dataset stores exhaustive top-15 retrieval from the 237,500 prompt–seed candidate audios for 48,840 background audios. One row per background contains 15 ranked candidates, including prompt, seed, clip ID, audio SHA256 and cosine score. Source audio stays in the two original public datasets; no audio is duplicated here.
This is an incremental run. Read progress.json before treating it as complete. The data/train split is an output packaging convention, not a new experimental split. run_config.json pins both input revisions, encoder and preprocessing. Embeddings and candidate index metadata are backed up under embeddings/ and index/. Results are published atomically with progress after each 1,000 backgrounds. The final manifest.json records hashes and counts for all result shards. This dataset contains retrieval results only, not generated or filtered audio.