Full fine-tuning (FFT)

Placeholder — write up FFT recipe details

What FFT is

A 3-stage pipeline:

1. Tokenizer replacement — swap Whisper's multilingual BPE for a per-lang BPE trained on monolingual text
2. Multitask fine-tuning — interleave ASR and text-only batches in a single training stage
3. Save best-by-criterion — Group A/B/C per language (see Save criteria)

When to use FFT

• Non-Latin scripts where Whisper's BPE fragments badly (chars/token < 2)
• Low-resource languages where text-only data is more abundant than ASR data

See also

• Simple fine-tuning (SFT)
• Save criteria
• Tokenizers → Recipe