← All blog posts

Project DoLittle

Humpback whale communication research, public datasets, and model releases.

A Careful First Pass at an Orca Acoustic Corpus

26,335 provenance-backed Orca contexts, source-aware quality triage, and a conservative cleanup path for future tokenization evaluation.

Hybrid LLM Architectures for Long-Context Whale Audio

How Hybrid SWA MoE and DeepSeek-style Hybrid NSA MoE architectures extend audio context while preserving a causal transformer decoder.

Public Release of Trained Neural Audio Models

Three trained DAC 9-codebook checkpoints for humpback whale communication — 10K, 32K, and 128K models for inference and fine-tuning.

Public Dataset Release: 10B Tokens of Humpback Whale Song

10 billion tokens of humpback whale song from NOAA SanctSound recordings, tokenized with 9-codebook DAC for long-context audio research.

MMC Humpback DAC9 Model Release Details

Technical details for the released humpback whale DAC9 model checkpoints.