Installation

Placeholder — fill in as code-release plan firms up

This page should cover environment setup once the code-release path is decided.

Hardware requirements

• GPU: 1× A100 80GB or H100 (FFT runs require ~38 GB VRAM for Whisper-large-v3 at batch_size=2, grad_accum=16)
• Disk: ~6 GB per checkpoint × ~300 ckpts = ~2 TB for the full matrix
• RAM: ≥64 GB recommended (goldfish corpora can be 10 GB)

Python environment

# Python 3.10+
python -m venv .venv && source .venv/bin/activate

# Core deps
pip install \
    torch \
    transformers \
    datasets \
    tokenizers \
    jiwer \
    librosa \
    soundfile \
    regex \
    wandb

External data

• FLEURS: pulled from google/fleurs via HuggingFace datasets
• CommonVoice v25: download from https://commonvoice.mozilla.org/datasets
• Goldfish text corpora: download from https://github.com/tylerachang/goldfish

Repository checkout

(To fill in once the code is released publicly.)

See also

• Single FFT job — for running a single fine-tune end-to-end
• Reproducing paper results — for the full pipeline