docs/SCHEMA.md

# Planned Parquet schema

Schema version: `bagpiper-self-contained-parquet-v1`.

Each row is one complete training sequence and contains both text and encoded
audio. Files are split into deterministic Parquet shards targeting about
256 MiB each.

| Field | Type | Description |
|---|---|---|
| `row_id` | string | Globally unique `source_subset:example_id` release key |
| `example_id` | string | Original source identifier (not globally unique) |
| `partition` | string | `generation` or `understanding` |
| `source_subset` | string | Internal source group retained for provenance |
| `split` | string | Release split, initially `train` |
| `direction` | string | `text_to_audio` or `audio_to_text` |
| `messages` | list of structs | Ordered `{role, modality, content, audio_index}` messages |
| `audio` | struct | `{path, bytes}` for the complete encoded target/input audio in this row |
| `audio_sha256` | string | SHA-256 of the encoded bytes |
| `audio_bytes` | int64 | Encoded byte length |
| `audio_format` | string | Container/codec suffix reported by the source file |
| `sample_rate` | int32 | Decoded sample rate, where readable |
| `channels` | int16 | Channel count |
| `num_frames` | int64 | Decoded frame count |
| `duration_seconds` | float64 | Decoded duration |
| `metadata_json` | string | Sanitized source annotations; no paths or usernames |
| `source_revision` | string | Pinned raw Hugging Face source revision |
| `schema_version` | string | `bagpiper-self-contained-parquet-v1` |

For an audio message, `content` is a sanitized display filename and
`audio_index` points to the embedded `audio` value. Text-message content is
preserved verbatim. Private source paths and usernames are never stored.

Every selected sequence has exactly one audio object. Intended reuse of the
same valid audio across multiple rows is preserved.

The audit found source-local ID collisions, so consumers should use `row_id`
as the key. `example_id` is preserved for traceability.