luvox-ai/voice-bench

Tool-call evaluation request context

tool_vi and halucinate contain 80 samples each. Both use the same realistic Vietnamese smart-home assistant context and complete mono 16 kHz WAV audio. tool_vi uses seeded random WHAM noise targeting 25–50% of speech RMS (approximately 6–12 dB SNR), capped at 50% over both the speech interval and the full background track. The cap may reduce noise further for uneven tracks. halucinate stays noise-only.

The dataset viewer contains only these eight columns:

Column	Purpose
id	Stable sample identifier.
audio	Complete input audio.
prompt	Text-modality input; empty for noise-only halucinate.
system_prompt	Actual assistant instructions and context.
tools	JSON string containing the complete tool schemas.
expected_tool	Expected tool name, or null for abstention.
expected_args	Expected arguments as JSON, or null.
default_city	Dataset context for scoring default-city weather requests.

The inference runner reads tools and system prompts from each sample. It must not insert fixed tool schemas or a default system prompt. Decode JSON strings before use:

import json
from datasets import load_dataset

samples = load_dataset("luvox-ai/voice-bench", "tool_vi", split="test")
sample = samples[0]
tools = json.loads(sample["tools"])
system_prompt = sample["system_prompt"]

Generation metadata (noise sources, TTS, SNR, durations, checksums) and duplicate labels such as expected_tool_calls remain in provenance/<config>/metadata.jsonl and the accompanying manifests. The compact viewer preserves all audio, request context and scoring behavior. Other dataset configurations are preserved.

Home tool v1: single-turn Lumi smart-home evals

These versioned configs are additive; the existing tool_vi, tool_en, and tool_pw733 configs remain available and are not overwritten by this release.

Config	Content	Scoring
home_single_vi_v1	400 Vietnamese Edge TTS commands over four Lumi-style home inventories	Exact tool-call count, name and arguments
home_single_en_v1	400 English one-turn commands; canonical inventory names may stay localized	Exact tool-call count, name and arguments
home_single_pw733_v1	126 original PW733 recordings with Lumi-style context and tools	Exact tool-call count, name and arguments

Every sample is one user utterance. A turn may produce multiple native tool calls: same-feature groups use one call with multiple device_ids; distinct actions use separate calls in the same turn. System prompts are in English, with device, room, floor and scene names retained from the supplied home context. Device control uses device_ids arrays for both one-device and multi-device requests. The eval scores calls but does not execute home actions. Created 2026-10-06, this version adds Lumi-aligned home inventories and capabilities, English system prompts, one-turn coverage, and the reviewed real-audio PW733 variant while leaving the earlier tool configs unchanged.

Text-only verification (--modality text --multimodal-schema luvox) against the two available servers:

Config	lumi-omni-14B (127.0.0.1:28080)	Qwen3.8-27B (10.10.0.116:8001)
home_single_vi_v1	61.25% (245/400)	83.00% (332/400)
home_single_en_v1	68.25% (273/400)	89.25% (357/400)
home_single_pw733_v1	62.70% (79/126)	88.89% (112/126)

The first endpoint returned lumi-omni-14B from /v1/models; it did not identify as the previously expected Qwen3.5-9B. Neither endpoint reached 95%. Scores are diagnostic and were not used to change the reviewed samples.

Audio verification is reference-only and uses --modality audio. The 127.0.0.1 endpoint used --multimodal-schema luvox with Luvox-3-flash-0930; the 10.10.0.116 endpoint used the standard OpenAI audio schema with run label Luvox-3-1006. Its API rejected that alias as the model field, so requests used the ID advertised by /v1/models.

The first PW733 build had a label-generation defect: some group_control labels contained empty device_ids, and several no-tool/status labels differed from the reviewed candidate. The config was corrected from the reviewed transcripts; the table below reports scores recalculated against those labels.

Config	Luvox-3-flash-0930 (127.0.0.1:28080)	Luvox-3-1006 (10.10.0.116:8001)
home_single_vi_v1	45.00% (180/400)	79.25% (317/400)
home_single_en_v1	42.50% (170/400)	75.00% (300/400)
home_single_pw733_v1	34.92% (44/126)	58.73% (74/126)