OpenVoiceOS/ovos-intents

OVOS intents

This is the canonical intent corpus of the OVOS skill fleet. Each row is one utterance with the intent label it belongs to. The train split comes from each skill's .intent and .entity resources at each skill's dev head. The test split is the skills' own end-to-end golden utterances. Labels have the form <skill_id>:<intent_name>, as OVOS-INTENT-4 defines.

Both splits ship here. The version of the content is the tag: v6, v6.1, v7-preview and v7-preview2 exist, and the repository name carries no version. Pin a tag to get one build of both splits:

from datasets import load_dataset

ds = load_dataset("OpenVoiceOS/ovos-intents", revision="v7-preview2")

v7-preview2 is a preview, not a release. It is built from every skill's dev head rather than from pinned commits, and the gaps it is known to carry are listed under "Known not yet fixed". No accuracy claim about a model trained on this corpus is supported until a golden evaluation against it is published.

Supersedes

These datasets are retired and stay available for reproducibility:

• OpenVoiceOS/ovos-intents-train-v1
• OpenVoiceOS/ovos-intents-train-latest
• OpenVoiceOS/ovos-intents-v5-eval
• OpenVoiceOS/ovos-golden-utterances-bench
• OpenVoiceOS/OVOSGitLocalize-Intents
• OpenVoiceOS/MT-intents-dataset-pt-PT
• OpenVoiceOS/ovos-llm-augmented-intents
• OpenVoiceOS/ovos-weather-intents
• OpenVoiceOS/ovos-common-query-intents
• OpenVoiceOS/ovos-intents-massive-subset
• OpenVoiceOS/ovos-intents-ilenia-testset-ca
• OpenVoiceOS/ovos-intents-ilenia-testset-es
• OpenVoiceOS/ovos-intents-ilenia-testset-nl

Layout

One directory per locale:

{lang}/train_templates.jsonl    the training rows       53 locales
{lang}/test.jsonl               the gold rows           51 locales

The file pattern is the same as in v6.1, v6 and v5, so a consumer that reads one of these paths keeps working when it moves to a new tag.

How the splits are made

The train side is every sentence a skill's templates expand to, with slots filled from the skill's own entity values in the same language. Typed slots (numbers, dates, durations, colours, languages) are filled by the OVOS parsers for that language. A language with no parser support for a type has no rows for templates that need it.

The test side is the gold sentences each skill ships for its end-to-end tests. The gold side is never thinned.

The two sides are kept disjoint sentence by sentence. Every training row whose utterance equals a gold utterance, after lowercasing, collapsing whitespace and stripping leading and trailing punctuation, is removed from the train side. That removal took 1,913 rows. Afterwards both checks report zero overlap:

Check	Rows in both splits
plain string comparison	0
comparison ignoring punctuation	0

When every expansion of a template collides with gold, the whole template leaves the train side. 343 templates are silenced this way. A silenced template is a sentence shape the model meets only in evaluation. Each one is listed in manifest.json.

How it was built

• Builder: train/build_from_skills.py in OpenVoiceOS/ovos-m2v-pipeline, pull request #329 (head 5030e87) on dev 4030dee, with the corpus extra installed so every typed-slot parser is present.
• Every source repository is read at its dev head, not at a pin. The commit each one resolved to is recorded under source_refs in manifest.json, so the build can be repeated exactly.
• The label function is train/skill_labels.py, with no spelling fallback.
• Every build gate passed: no typed-slot placeholder or resource path left in a row, no typed-slot template without rows that a language gap does not explain, zero train and gold overlap, and no label lost that the ledger does not list. The ledger lists no pending removal or addition.
• The per-locale files hold the builder's rows byte for byte. Row count and an order-independent hash of the rows agree between the builder's output and this tree, for both splits.

Fields

Field	Meaning
lang	BCP 47 language tag of the utterance, for example ca-ES
label	<skill id>:<intent name>, for example ovos-skill-weather.openvoiceos:weather
utterance	the sentence
source	skill:<repo> for train rows, gold:<repo> for test rows
template	the template the sentence was expanded from (train rows only)
needs_manual	test rows only: the skill's gold file marks the row for a human check. Whether that check happened is unrecorded. These rows are ordinary gold rows and are scored like the others

Counts

Count	v6.1	v7-preview2
train rows	2,236,404	4,952,006
train locales	52	53
gold rows	2,380	12,908
gold rows marked needs_manual	not recorded	10,496
gold locales	21	51
labels trained	232	221
labels scored	191	203
labels never scored	41	18

The new train locale is es-CO. Every v6.1 locale is still present.

Twenty-three labels of v6.1 are gone and twelve are new. The twelve new labels are renames to OVOS-INTENT-2 base names (cmd, color-picker, laugh, naptime, volume). The other eleven losses are intents the skills folded into other labels or removed. The ledger in train/labels.yaml gives the reason for each and, where there is one, the label its phrasings went to.

The six largest train locales are en-US 737,362, ca-ES 480,397, pt-BR 405,642, da-DK 291,481, pl-PL 291,478 and fr-FR 281,268. 27 locales hold fewer than 100 rows each.

Gold rows per locale: en-US 1,506, pt-BR 655, nl-NL 636, da-DK 629, pt-PT 623, it-IT 619, sv-SE 617, es-ES 615, eu-ES 615, de-DE 614, ca-ES 608, gl-ES 608, fr-FR 601, ru-RU 540, kab 521, pl-PL 502, es-CO 491, fa-IR 487, oc-FR 249, cs-CZ 211, hu-HU 197, fi-FI 149, sv-FI 146, tr-TR 118, el-GR 99, an 45, ro-RO 38, lt-LT 19; 9 each in bg-BG, he-IL, hr-HR, id-ID, ja-JP, nb-NO, sk-SK, sl-SI and th-TH; 6 each in ar-XA, et-EE, fil-PH, ko-KR, lv-LV, ms-MY, uk-UA, vi-VN and zh-CN; 3 each in ar-XX, cmn-CN, is-IS, nn-NO and zsm-MY.

Two locales have no gold

es-LM and eu have training rows and no test.jsonl at all. No empty file is shipped for them, because an empty file reads as "measured, found nothing" when the truth is "never measured". 18 of the 221 labels are never scored.

Known not yet fixed

1. Place names are not declined in pl-PL, ru-RU, ca-ES and gl-ES. A location slot is filled with the nominative form, so rows read "w Norwegia" and "w Izrael" where Polish needs "w Norwegii" and "w Izraelu".

Caveats

1. The label sizes differ by a factor of about 94,000. date-time what_time_will_it_be has 937,379 rows (18.9% of train), pokepedia battle 744,000 (15.0%) and weather weather 482,629 (9.7%), while the smallest, randomness make_a_choice, has 10. Cap rows per label before training; train/train.py --max-per-label in the pipeline does this with stratification by language.
2. 33 noncompliant base names remain. Their labels change when their skills rename.
3. 57,097 rows keep an unfilled slot, a {slot} the build had no example value for. The runtime leaves such a slot as free-form capture, so the placeholder is kept literal.
4. 2,830 rows are ambiguous: each shares its locale and utterance with a row carrying a different label.
5. The machine_generated flag in the skills' gold files is unreliable. Every gold sentence in this fleet was written by a model, so the flag is wrong wherever it claims False. The builder does not read it and the test rows here do not carry it.

Files

• {lang}/train_templates.jsonl: the training rows for that locale.
• {lang}/test.jsonl: the gold rows for that locale, where gold exists.
• manifest.json: the full build record, including the silenced templates, the label ledger and the per-stage counts.

Credit

Funded by the NGI0 Commons Fund / NLnet under grant agreement No 101135429, through the European Commission's Next Generation Internet programme.