luvox-ai/voice-reply

Voice Reply

150 Vietnamese text cases for spoken replies after tool results and clarification. Synthetic homes; no tools are executed.

Config	Groups	Samples
status	large, room, single	30
control	partial, pending, success	30
repair	loop, missing	20
failure	errors	10
news	details, headlines	20
wiki	knowledge	10
weather	forecast	10
spoken	normalization	10
safety	untrusted	10

Every group has 10 cases. Split: test only. Five contexts: small (8 devices), family (20), large (42), ambiguous (20), limited (8).

Input: replace ${context} in prompt-A.txt or prompt-B.txt with context; send parsed messages and tools. JSON-valued columns are strings. Tool replies retain the nested data.res JSON string; plain-text errors are preserved. Gemini requires native function-response conversion; Luvox-2 uses its raw tool-call adapter.

Judge: pass user input, relevant history/tool result, ground_truth, grading_facts and candidate output to the common rubric.json. Do not send the full home context, model name or A/B identity. One JEV rating, 1–5. Average over graded cases; report coverage/errors separately. Not a speech or TTS listening benchmark.

Sources: 20 real web snapshots: 10 publisher news abstracts and 10 Vietnamese Wikipedia revision extracts. URLs, dates and attribution are in sources; snapshot facts are not live current facts or independently verified news. Other cases are synthetic.

Rights: original fixtures and annotations: CC BY 4.0, Luvox AI. Wikipedia excerpts: CC BY-SA 4.0, Wikipedia contributors, linked revisions. News titles/short abstracts retain publisher rights and attribution; no full articles are distributed. Source terms govern reused content.

Status: v1 pilot/review dataset. Prompts A/B are candidates; no claim that B is better. Model outputs are stored separately from the dataset.

Evaluation v2: rubric.json and judge-facts.json clarify equivalent curtain/color/fan states, partial results and spoken UX. Use the relevant entry from judge-facts.json as semantic_facts in the judge state. Model inputs, 150 cases and A/B prompts are unchanged. Prior rubric is retained in rubrics/voice-reply-v1.json; compare scores within the same rubric version.