soda-research/soda-600m-prelim

🥤 SODA Preliminary Run

SODA (Scaling Open Discrete Audio) is a suite of discrete audio foundation models using next-token prediction on interleaved semantic, acoustic, and text tokens.

🌐 Project Page: https://soda-audio.github.io

This model is part of our SODA Research collection (e.g., IsoFLOP sweep, preliminary runs, or ablations).

Note that this model (soda-600m-prelim) is the same model as in https://huggingface.co/WillHeld/blueberry

As an preliminary experiment, we were not strict with our data criterion, and we ended up training this model on multiple languages (%ratio) as follows: including English (72.52%), Spanish (15.04%), French (7.13%), German (4.73%), Thai (0.20%), Hindi (0.16%), Arabic (0.12%), Chinese (0.10%). This model was trained on 500B tokens in total.

For full usage instructions, and more information, please refer to the SODA-4B-base model card:
👉 SODA-4B-base

📈 WandB: https://wandb.ai/marin-community/marin/runs/exp1699_marin_yodas2-b5edae/