OpenVoiceOS/ovos-golden-utterances-bench

OVOS Golden Utterances Benchmark

A held-out evaluation set of hand-curated golden utterances drawn from the end-to-end test suites of the OpenVoiceOS skill fleet. Each row pairs an utterance with the skill intent it is expected to trigger. The utterances were written independently of the skills' own intent training templates (vocab/intent files used to train the classifiers), so the set is suited to measuring real intent-matching accuracy rather than re-testing the training data itself.

Schema

One JSONL file per language at <lang>/test.jsonl. Columns:

Column	Type	Description
utterance	string	The natural-language phrase a user would say.
expected_intent	string	<skill_id>:<intent_label>, the intent the utterance must resolve to.
lang	string	BCP-47 locale of the utterance (e.g. en-US, pt-PT).
skill_id	string	The OVOS skill id that owns the intent.
source_repo	string	The ovos-skill-* GitHub repository the row was extracted from.
source_file	string	Path to the source golden-utterances file within that repo.
intent_label_original	string	The intent label exactly as it appears in the source suite, before normalization.
intent_type	string or null	The intent engine the utterance targets (e.g. padatious, adapt).
intent_method	string or null	The skill's handler method for the intent, when recorded.
needs_manual	bool	Whether the row is flagged as requiring manual review rather than automated scoring.
machine_generated	bool or null	Whether the utterance was machine-generated. null when the source row did not record this.
required_vocab	list or null	Vocabulary terms the utterance is expected to require, when recorded.
expected_messages	list or null	Bus message types the skill is expected to emit in response, when recorded.

Intent label normalization

Source suites label file-based (Padatious) intents with a trailing .intent suffix, matching the .intent training file name, while runtime dispatch strips that suffix before emitting the intent on the bus. expected_intent always uses the stripped form, composed as <skill_id>:<intent_label without .intent>. Adapt-era intents, which never carried the suffix, compose unchanged. The original, unstripped label is kept in intent_label_original for provenance.

Rows without a lang field in the source suite are the fleet's default, en-US. Some suites ship as a single default-locale file per skill; others ship one file per locale, named golden_utterances_<locale>.jsonl — in that case every row in the file is stamped with that file's locale, unless the row already carries its own lang, which wins.

Flagged rows

Rows with needs_manual: true or machine_generated: true are included in the dataset, not excluded, so that per-skill and per-language totals stay complete. Consumers who want a strictly human-curated, auto-scorable subset should filter both fields to false.

Per-language row counts

Language	Rows
ca-ES	33
cs-CZ	15
da-DK	34
de-DE	34
en-US	683
es-ES	33
eu-ES	34
fa-IR	8
fr-FR	34
gl-ES	34
hu-HU	15
it-IT	34
nl-NL	25
pl-PL	15
pt-BR	11
pt-PT	34
ru-RU	15
sv-FI	15
sv-SE	15
Total	1121

License

Apache-2.0, matching the license of the source ovos-skill-* repositories.