Orvena/Cue-0.4B-data

Cue-0.4B data

This dataset contains the text used to train and test Cue-0.4B, the turn-taking model in Orvena's voice mode. Cue makes two decisions from the speech recognizer's transcript and the assistant's own words. When the user speaks while the assistant is talking, it decides whether to keep going, stop, or treat the speech as a new request, and where that request starts. When a new turn begins, it decides whether the turn needs the main language model or is one of nine quick intents the assistant can answer at once: presence check, greeting, thanks, acknowledgement, cancel, wait, repeat, goodbye and "how are you".

The data covers 30 languages. All of it was written by language-model agents for this purpose; none of it comes from real users or recorded conversations.

Contents

Path	Size	Description
interruptions/banks/<lang>.json	30 files	Interruption phrase banks, 302 entries and 24 development examples per language
new-turns/banks/<lang>.json	30 files	New-turn phrase banks, about 650 entries per language
composed/train.jsonl, validation.jsonl, test.jsonl	63,000 / 9,000 / 9,000 rows	Examples composed from the banks, as the model saw them
interruptions/dev.jsonl	720 rows	Interruption development set (the banks' natural_test items)
interruptions/test.jsonl	900 rows	Interruption test set
new-turns/test.jsonl	1,800 rows	New-turn test set
results/interruptions-test.jsonl, new-turns-test.jsonl	900 / 1,800 rows	Outputs of the released model on the two test sets

Phrase banks

Every interruption bank has the same keys: orvena (40 sentences the assistant says), backchannel (24), filler (8), keep_going (12, plus 8 in keep_going_extra), reaction (20), side_talk (16, plus 10 in side_talk_extra), background (16, speech from a television or other people), stop (30), stop_long (20, stops with a reason), request (60, plus 30 in request_extra), connector (8, words such as "now" and "and") and natural_test (24 complete examples). The *_extra lists are merged into their base list when examples are composed.

{"language": "en",
 "orvena": ["Right now it's 22 degrees and sunny in Austin, with a light breeze from the south and humidity around 45 percent.", "..."],
 "backchannel": ["mm-hmm", "uh-huh", "..."],
 "stop": ["stop", "wait", "..."],
 "request": ["will it rain tomorrow morning?", "..."],
 "natural_test": [{"orvena": "...", "spoken_until": "...", "user": "Mhm.", "action": "continue", "span": null}, "..."]}

A new-turn bank has intents (phrases for each quick intent, keyed presence, greeting, thanks, ack, cancel, wait, repeat, goodbye, how_are_you), agent_lookalikes (phrases that share words with a quick intent but need the main model, such as "thanks, now add milk to the shopping list"), answers (replies to an assistant question, which also go to the main model), orvena_questions (assistant questions used as context) and replies (the fixed lines Orvena says for each intent, not used in training). Across languages the banks hold 12,907 intent phrases, 4,291 look-alikes, 1,413 answers and 968 questions.

{"language": "en",
 "intents": {"presence": ["can you hear me?", "are you there?", "..."], "thanks": ["thanks", "thank you", "..."]},
 "agent_lookalikes": ["hi, what's the weather like tomorrow?", "..."],
 "answers": ["yes", "yes please", "yeah, do it", "..."],
 "orvena_questions": ["Should I add it to your calendar?", "..."],
 "replies": {"presence": ["Yes, I can hear you.", "..."]}}

Test and development sets

In interruptions/test.jsonl and interruptions/dev.jsonl, orvena is the assistant's sentence and spoken_until the part already spoken when the user started (the whole sentence if the user spoke after it ended). user is the transcript, action is continue (keep going), stop or request, and span is the request, a suffix of user, or null.

{"id": "en-024", "lang": "en", "orvena": "The hardware store on Fifth Avenue closes at six today and opens again tomorrow at seven in the morning.", "spoken_until": "The hardware store on Fifth Avenue closes at six today and opens again tomorrow at seven in the morning.", "user": "Okay. And do they sell paint?", "action": "request", "span": "And do they sell paint?"}

In new-turns/test.jsonl, orvena is what the assistant said last (empty at the start of a conversation) and intent is agent (the main model) or a quick intent. After a question, an answer such as "sure thanks" is agent.

{"id": "en-047", "lang": "en", "orvena": "Should I remind you again in 30 minutes?", "user": "sure thanks", "intent": "agent"}

Composed examples

The files in composed/ hold the prompts exactly as they were used for training, calibration and the synthetic test. task is barge for interruptions and turn for new turns. action is 0 (keep going), 1 (stop) or 2 (request); intent is 0 (main model) or 1 to 9 for presence, greeting, thanks, ack, cancel, wait, repeat, goodbye and how_are_you; -100 marks a label that does not apply to the row. user_start and user_end are character offsets of the user's words in text, span_start is the offset where a request starts (-1 if there is none), and category is the composition category.

{"lang": "en", "category": "ack_request", "task": "barge", "action": 2, "intent": -100, "text": "Orvena speaking: … hours and 52 minutes, with about an hour and 10 minutes of deep sleep, and you | woke up twice.\nUser: sure, and wake me up at 6:30 tomorrow\nDecision:", "user_start": 121, "user_end": 158, "span_start": 127}

Results

The files in results/ have one row per test example, joined to the test files by id. They hold the temperature-scaled probabilities of the released int8 Core ML model, run with the app's prompts. For interruptions, label is 0, 1 or 2 as above, speaking is true when the user started inside the sentence, probabilities are keep going, stop and request, and pointer_best, pointer_safe and pointer_label are token positions in the prompt: the most likely request start, the earliest start with probability at least 0.15, and the labelled start (-100 when the example is not a request). For new turns, label and the order of probabilities follow the model's intent names (agent, presence, greeting, thanks, acknowledgement, cancel, wait, repeat, goodbye, howAreYou), and after_question marks examples whose context ends in a question.

{"id": "en-001", "lang": "en", "label": 0, "speaking": true, "probabilities": [0.989519, 0.00437, 0.00611], "pointer_best": 32, "pointer_safe": 32, "pointer_label": -100}

How training examples are composed

Each bank list is split by item into 70% training, 15% validation and 15% test with a fixed seed per language, so no test phrase appears in training; fillers and connectors are shared by all splits. An interruption example draws a category with the weights below, then an assistant sentence. A sentence longer than three words is cut at a random word with probability 0.85 (speaking, with | at the cut); otherwise it is complete (finished). The user's words are assembled from the category's pieces, and 5% of requests in languages other than English are English requests.

Action	Categories and sampling weights (%)
keep going	backchannel 16, filler 3, keep going 5, reaction 6, side talk 6, background 5, echo of the assistant's words 5
stop	stop 11, long stop 7, backchannel or filler then stop 4
request	request 12, acknowledgement then request 8, filler then request 4, stop then request 5, echo then request 2, side talk then request 1

A new-turn example is a quick phrase (44%), a pair of quick phrases with a fixed resulting intent (6%), a look-alike (22%), a plain request (10%), a request after a greeting, acknowledgement, thanks, presence check or cancel (8%), an answer after an assistant question (6%) or a quick phrase after an assistant question (4%). The context is empty 35% of the time, an assistant statement 45% and an assistant question 20%. After a question, acknowledgements, thanks, greetings and "how are you" are labelled agent.

To imitate speech recognizer output, 40% of examples lose their punctuation and, except in Japanese, their case, and multi-word pieces repeat a word with probability 0.05. Each language contributes 1,200 interruption and 900 new-turn examples to training and 150 of each to validation and to the synthetic test. The training script dropped 4 training rows that exceeded 128 tokens or whose request start was not on a word start, so the model was trained on 62,996 examples.

Test sets

Both test sets were written separately from the training data, in all 30 languages, by writers who did not see the phrase banks. The interruption test set has 30 examples per language: 300 keep going, 240 stop and 360 request, and in 609 of the 900 the user starts before the sentence ends. The new-turn test set has 60 examples per language, 33 for the main model and 3 for each quick intent; 822 of the 1,800 follow an assistant sentence. The development set (240 keep going, 180 stop, 300 request) was written together with the banks and used to choose the keep-talking threshold, so it is not independent of them.

Results

Cue-0.4B as the released int8 Core ML package, with the shipped thresholds (keep going only when its probability is at least 0.7, a quick reply only at 0.9 or above). The numbers can be recomputed from results/.

Interruptions (900 examples)
Right decision	90.4% (814)
Stopped when the user only acknowledged	11.0% (33 of 300)
Kept talking through a stop	10.8% (26 of 240)
Missed a request	4.4% (16 of 360)
Request start on the labelled word, for requests decided as requests	92.6% (315 of 340)

New turns (1,800 examples)
Right decision	95.2% (1,713)
Requests for the main model answered as a quick intent	1 of 990
Quick intents sent to the main model	10.4% (84 of 810)
Quick intent answered with the wrong one	2 of 810

Notes

• The prompts use the assistant name "Orvena".
• The data is written text. Recognizer errors are imitated, not recorded.
• Some sentences mention public figures, businesses and places, as in music requests, sports commentary and general-knowledge questions. The people in calendar, message and contact examples are fictional.
• Before release, phone numbers spelled out in a few assistant sentences of the development and test sets were replaced with obviously fictional ones. The results for the four affected test examples (es-022, hi-023, hu-023, lv-022) were computed on the original wording.

License

The data is released under CC BY 4.0. The model is released separately under Apache-2.0.

Citation

@misc{orvena2026cue,
  title        = {Cue: Multilingual Turn-Taking for On-Device Voice Assistants},
  author       = {{Orvena}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Orvena/Cue-0.4B}},
  note         = {Technical report: https://huggingface.co/Orvena/Cue-0.4B/blob/main/Cue-0.4B-technical-report.pdf}
}