---
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata/train.parquet
  - split: valid
    path: metadata/valid.parquet
  - split: test
    path: metadata/test.parquet
---

# duet

The table above is one row of metadata per row of the corpus. The rows themselves are
the tar shards in this repository, which carry the four channels as arrays and are
read by unpacking a snapshot rather than through a `datasets` builder.

| split | rows | shards |
| --- | ---: | --- |
| `train` | 20,808 | `train-*.tar` |
| `valid` | 11,460 | `valid-*.tar` |
| `test` | 9,660 | `test-*.tar` |