Deterministic synthetic data for predicting Piper semantic lexical units from Qwen3 assistant output. It contains Vietnamese and English utterances with the 14 kinds defined by the companion project. Each JSONL row has source_text, auditable Unicode start/end spans, and semantic units. Every split also includes natural Vietnamese/English code-switch examples, including held-out foreign clauses; lexical language is annotated per unit. Punctuation uses the host-sentence language canonically and every current production unit has stress none (no synthetic prominence claim).

Files

train.jsonl, validation.jsonl, test.jsonl: immutable split rows.
manifest.json: generator seed, counts, template IDs and SHA-256 values.

The spans are training/audit metadata only; downstream Piper requests emit only {text, language, kind, stress}. Whitespace is never a unit.

Reproducibility

Training downloads the published split files and validates their manifest/checksums. To change templates or values, materialize, review and upload a new dataset artifact; do not regenerate inside training.

Downloads last month: 82

Total file size:

14.2 MB