elizaos--eliza/packages/benchmarks/mmau-audio/AGENTS.md

MMAU Audio — Agent Guide

Audio MMAU (Sakshi et al., ICLR 2025): 10,000 audio clips across speech, sound, and music domains, 27 reasoning skills, all multiple-choice. Scoring is deterministic exact match — no LLM-judge required. Registered in the suite as mmau.

Run

# Direct, from this directory
python -m elizaos_mmau_audio --agent eliza --split test-mini --output ./results --json

# Subset by category
python -m elizaos_mmau_audio --agent eliza --split test-mini \
    --category speech --limit 100 --output ./results

# Through the suite orchestrator (resolves provider/model, stores results)
python -m benchmarks.orchestrator run --benchmarks mmau --provider eliza --model <m>

Smoke test (no API keys)

python -m elizaos_mmau_audio --mock --limit 2

--mock uses the bundled fixtures/smoke.jsonl (8 samples, all categories) and the oracle agent. Zero credentials required.

Test the harness

pip install -e .[dev]
pytest tests/ -x

Layout

Path	Role
elizaos_mmau_audio/cli.py	argparse CLI; also __main__.py entry
elizaos_mmau_audio/runner.py	Load → dispatch → score → persist
elizaos_mmau_audio/agent.py	OracleMMAUAgent, CascadedSTTAgent, AgentFn type
elizaos_mmau_audio/dataset.py	Bundled fixture + Hugging Face streaming loader
elizaos_mmau_audio/evaluator.py	Deterministic MCQ scoring + per-skill aggregation
elizaos_mmau_audio/types.py	MMAUSample, MMAUConfig, MMAUReport, enums
fixtures/smoke.jsonl	8-sample offline fixture (all 3 categories)
tests/	pytest suite (evaluator, dataset, runner)

Notes

• Results write to benchmark_results/mmau/<timestamp>/ (gitignored). The orchestrator result file is mmau-results.json.
• Scored by _score_from_mmau_json in registry/scores.py (line 805).
• Real runs (--agent eliza|hermes|openclaw) stream audio from Hugging Face (gamma-lab-umd/MMAU-test-mini 1k or gamma-lab-umd/MMAU-test 9k) and use Groq Whisper as a cascaded STT step (needs GROQ_API_KEY). The sound and music categories are lossy under this pipeline; treat them as a floor.
• AgentFn in agent.py receives raw audio_bytes, so a future direct-audio adapter can bypass the STT step.
• 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.