Hinglish 60k Mimi Q8 Style Tokens

TinyAya codec-tokenized style finetuning dataset built from /mnt/data/raw/hinglish-60k/metadata.jsonl and /mnt/data/raw/hinglish-60k/audio.

Rows are Mimi codec-only examples. The text field intentionally contains no speaker prefix; this artifact is meant to teach tones and inline emotion/event tags on top of a speaker-conditioned base model.

Tones: [happy] [angry] [neutral] [sad] [whisper] [excited]

Inline emotions/events: <laugh> <chuckle> <sigh>

No speaker_id field is emitted. The original source speaker id is preserved as source_speaker_id only so training cannot accidentally use this single-speaker dataset as a same-speaker reference pool.