OpenVoiceOS/massive-templates

MASSIVE Templates

A mechanical re-cast of AmazonScience/massive into the same row shape as OpenVoiceOS/intents-for-eval. Every MASSIVE utterance becomes one training row with a bracket template ({slot} placeholders) plus a list of slot examples drawn only from that single utterance. No LLM — the conversion is a single regex pass over MASSIVE's annot_utt field.

Source	AmazonScience/massive (Apache-2.0)
Languages	52 BCP-47 locales
Train rows	~13.5k per locale (train + dev pooled) — ~704k total
Test rows	2,974 per locale — ~155k total
Schema fit	Drop-in compatible with OpenVoiceOS/intents-for-eval consumers

How the conversion works

MASSIVE annotates utterances inline:

annot_utt = "wake me up at [time : five am] [date : this week]"

A single regex \[\s*([A-Za-z0-9_]+)\s*:\s*([^\[\]]+?)\s*\] extracts each [slot : value] span; the span gets replaced with {slot} in the template and (slot, value) is recorded for the row's slot list:

{
  "intent_id": "alarm:alarm_set",
  "domain":    "alarm",
  "template":  "wake me up at {time} on {date}",
  "slots": [
    {"name": "time", "examples": ["nine am"]},
    {"name": "date", "examples": ["friday"]}
  ]
}

The examples list carries only the values from that specific utterance — no cross-utterance aggregation. Two utterances with identical templates but different slot fills become two distinct rows.

Test rows mirror intents-for-eval's shape:

{
  "utterance":       "wake me up at five am",
  "expected_intent": "alarm:alarm_set",
  "expected_slots":  {"time": "five am"},
  "split":           "test"
}

Repo layout

<locale>/train_templates.jsonl
<locale>/test.jsonl

There is no train_keywords.jsonl because MASSIVE has no per-intent keyword pool to lift; engines that want one can derive it from the union of every template row's slot values at load time.

Reproducing

The converter and a ready-to-use loader live in OpenVoiceOS/ovos-intent-benchmark:

git clone https://github.com/OpenVoiceOS/ovos-intent-benchmark
cd ovos-intent-benchmark
pip install -e .
python scripts/convert_massive_to_templates.py --langs all --out data/massive_templates/

Loading

Via the benchmark loader (matches IntentsForEvalDataset):

from benchmark.datasets import get
ds = get("massive_templates")
train, test = ds.load("en-US")

Or directly via datasets:

from datasets import load_dataset
templates = load_dataset("OpenVoiceOS/massive-templates", "en-US-templates", split="train")
test = load_dataset("OpenVoiceOS/massive-templates", "en-US-test", split="test")

License

Apache-2.0, inherited from MASSIVE.

Citation

If you use this dataset, please cite the original MASSIVE paper:

@article{fitzgerald2022massive,
  title={MASSIVE: A 1M-Example Multilingual Natural Language Understanding Dataset
         with 51 Typologically-Diverse Languages},
  author={FitzGerald, Jack and Hench, Christopher and Peris, Charith and others},
  journal={arXiv preprint arXiv:2204.08582},
  year={2022}
}