elizaos--eliza/packages/benchmarks/voicebench-quality/CLAUDE.md

VoiceBench (quality) — Agent Guide

Vendored implementation of VoiceBench (Chen et al. 2024): 8 task suites covering 6,783 spoken instructions, measuring response quality (score in [0, 1]) for voice-input language assistants. Registered in the suite registry as voicebench_quality.

Separate from packages/benchmarks/voicebench/ (that one is TypeScript and measures latency in ms; this one is Python and measures response quality).

Run

# Direct, from this directory
python -m elizaos_voicebench \
    --agent eliza \
    --suite all \
    --output ./results

# Through the suite orchestrator
python -m benchmarks.orchestrator run --benchmarks voicebench_quality --provider <p> --model <m>

Agent choices: eliza, hermes, openclaw. STT provider is auto-detected (prefers local eliza1 binary > GROQ_API_KEY > faster-whisper); override with --stt-provider {groq,eliza-runtime,eliza1,faster-whisper,local-whisper}.

Smoke test (no API keys)

# --mock uses bundled fixtures with a deterministic no-cost adapter and judge
python -m elizaos_voicebench --mock --suite openbookqa --limit 5 --output /tmp/vbq-smoke

Note: mock results are rejected by the real scorer (_score_from_voicebench_quality_json).

Test the harness

pip install -e ".[test]"
pytest tests/ -x

Layout

Path	Role
elizaos_voicebench/__main__.py	CLI entry point (python -m elizaos_voicebench)
elizaos_voicebench/runner.py	Execution loop — resolves suites, drives adapter + judge
elizaos_voicebench/adapters.py	VoiceAdapter base + eliza/hermes/openclaw/echo impls
elizaos_voicebench/clients/	Groq STT, eliza-1 ASR, Cerebras LLM judge, say TTS
elizaos_voicebench/fixtures/	Bundled JSONL task fixtures (8 suites) used in mock mode
elizaos_voicebench/types.py	SUITES tuple and shared type definitions
tests/	pytest suite
pyproject.toml	Package metadata; elizaos-voicebench console script

Notes

• Results write to <output>/voicebench-quality-results.json. Registry expects it at <output_dir>/voicebench-quality-results.json (_voicebench_quality_result).
• Scored by _score_from_voicebench_quality_json in registry/scores.py (line 755). Score is the unweighted mean of the 8 per-suite scores.
• Required env vars for live runs: CEREBRAS_API_KEY (LLM judge), GROQ_API_KEY or VOICEBENCH_QUALITY_STT_PROVIDER (STT).
• Judged suites (alpacaeval, commoneval, sd-qa, bbh) use gpt-oss-120b on Cerebras; deterministic suites (ifeval, advbench, openbookqa, mmsu) need no judge key.
• Full background: README.md.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing in this package is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. This applies to every feature, fix, refactor, and chore here. "Tests pass" is not proof; "CI is green" is not proof.

• Record AND read model trajectories. Capture the actual inputs and outputs of the model from a live LLM — not the deterministic proxy, not a mock: the prompt, the providers/context, the raw model output, every tool/action call, and the result. Then open the trajectory and review it by hand. A captured-but-unread trajectory is not evidence (packages/scenario-runner/bin/eliza-scenarios run <scenario> --report <out>).
• Real, full-featured E2E — no larp. Every feature ships detailed end-to-end tests that drive the real path end to end. Not the happy "front door" only: cover error paths, edge/empty/invalid input, concurrency, roles/permissions, and adversarial input. A test that asserts against a mock/stub/fixture standing in for the thing under test does not count. If the real model/device/chain/connector/account is hard to reach, make it reachable — that is the work, not an excuse to mock. If the existing tests here are shallow or mocked, fixing them is part of your change.
• Screenshots + logs at every phase, plus a complete walkthrough video/run-through of the entire feature or view, start to finish (bun run test:e2e:record).
• Manually review every artifact the change touches — never just the green check: client logs (console + network), server logs ([ClassName] …), the model trajectories in and out, before/after full-page screenshots, and the domain artifacts listed below for this package.
• No residuals. No shortcuts. The goal is not "done" — it is everything done. Clear every blocker by the hard path: build the real architecture, stand up the real model/device/service, actually test it. Never leave a TODO, a stub, a stepping-stone, or a "follow-up." When unsure, research thoroughly, weigh the options, and ship the best, highest-effort, production-ready version. Keep going until every possibility is exhausted.

Artifacts → attached inline in the PR (MP4 video, JPG screenshots, logs in <details>); attach each evidence type or explicitly mark it N/A with a reason — never leave it blank. If develop moved and changed behavior, re-capture evidence; stale proof is worse than none.

Capture & manually review for this package — benchmark / eval suite:

• A real-model run (not the mock/smoke fixture) producing the score-report JSON, with the numbers inspected and the provider/model recorded.
• The per-item trajectories the harness captured, spot-reviewed for correctness — a green harness run over mock fixtures is not a result.
• The provider matrix actually exercised, and the scoring math validated against a known case.
• Failure / timeout / partial-output handling in the harness itself.