# s2s-a0-en — Stage-A0 training grids for a small duplex S2S model

Format per `tokens/a0_*.npz`:
- `audio` int16 [T,16] — Mimi codes @12.5Hz, 8 codebooks: cols 0-7 = speech
  ("agent" channel), cols 8-15 = encoded silence ("user" channel)
- `text` int32 [T] — Qwen3-0.6B token ids, run-length from each utterance's
  start frame; pad elsewhere
- `prompt_len` = 0, `source` = "a0_<dataset>"
Utterances packed to ~100s per grid with 0.2-0.4s gaps. No delay/lead applied.
`manifests/*.json` map each source parquet file to its grids + hours.
See ATTRIBUTION.md for source datasets and licenses.