cairninstitute/mmc-orca-dac9-cc-by-sa-4

MMC Orca DAC9 (CC-BY-SA-4.0)

License-compatible, tokenized Orca contexts published by CAIRN Institute for Project DoLittle.

Contents

This repository contains 8,560 evaluator-graded 30-second Orca contexts encoded as Descript DAC 44.1 kHz, nine-codebook integer arrays (198,994,320 token IDs). The arrays are stored in one uncompressed WebDataset TAR shard:

• —data/webdataset/mmc-orca-dac9-cc-by-sa-4-00000.tar - token arrays, grouped by upstream source identifier
• —metadata/manifest.csv - source provenance, license, evaluation grade, exact probabilities, preparation route, and per-array SHA-256
• —metadata/shard_manifest.csv - mapping from every context to its TAR member
• —metadata/release_manifest.json - release-level counts and codec configuration
• —metadata/SHA256SUMS - per-array checksums retained from the verified tokenization audit

The shard is intentionally uncompressed to support sequential streaming and to avoid publishing thousands of individual LFS objects. Raw recordings and denoised WAVs are not included.

Sources

• —dclde_dfo_crp: 3,116 contexts
• —dclde_dfo_wdlp: 1,710 contexts
• —dclde_orcasound: 259 contexts
• —dclde_scripps: 90 contexts
• —dclde_simres: 560 contexts
• —dclde_smru: 159 contexts
• —dclde_uaf: 502 contexts
• —dclde_vfpa: 2,164 contexts

Accessing The Tokens

Download the dataset snapshot, then extract the shard. Extraction preserves the data/<source_id>/ layout referenced by metadata/manifest.csv.

mkdir -p data/tokenized/orca_dac9
tar -xf data/webdataset/mmc-orca-dac9-cc-by-sa-4-00000.tar -C data/tokenized/orca_dac9

The shard can also be streamed with a WebDataset-compatible loader:

import webdataset as wds

dataset = wds.WebDataset("data/webdataset/mmc-orca-dac9-cc-by-sa-4-00000.tar")

Quality And Provenance

Only frozen-selection contexts with evaluator grades quiet, soft, or moderate are represented. The manifest retains the original grade, all exact evaluator probabilities, preparation threshold, source license, source manifest, recording order, and time bounds. Four processing holds and contexts whose source terms do not support this license partition are excluded from this repository.

License And Use

This repository is distributed under CC-BY-SA-4.0. Preserve source attribution, the accompanying license notice, and per-context provenance when redistributing. Do not combine this partition with token data under incompatible source licenses. CAIRN Institute does not claim rights beyond its contribution to the selection, organization, processing, and packaging of the release.