hyper07/watch-me-safety-scenes

Watch Me safety scenes

A small benchmark for one question: listening to a conversation around someone, should a safety assistant ask her "Are you OK?" It has spoken scenes with their scripts, labelled harmful or harmless, and written accounts of harassment at school, work and online, labelled with what an official source concluded.

It is the test set of Watch Me, a voice-locked safety assistant built for the Meta Global AI Developer Hackathon 2026, published so other systems can be measured the same way. Every voice is synthetic and every scene is fictional.

What's in it

Config	Rows	What
scenes	22	the scene's audio (16 kHz WAV), with its script line by line: who speaks (owner is the person being protected), the synthetic voice, the text, start and end in seconds, and harm_starts on the first harmful line. 11 harmful, 9 harmless hard cases, 2 school and work (Notice)
enrollment	3	the owner saying her enrollment line, for voice-lock (speaker verification) tests
notice_public	40	public anti-harassment training examples with an official answer (EEOC, New York State, StopBullying.gov), rewritten as first-person accounts; expect lists the kinds of conduct the source named, official its conclusion, with source and url
notice_ours	16	accounts written for Watch Me, 8 that should match and 8 that should not

Harmful scenes: drink pressure then "I feel dizzy"; unwanted touching; "don't tell anyone, nobody will believe you"; followed on the way home; others talking about her while she can't respond; quiet pressure on a date; street harassment; a stranger saying "stop watching"; a rideshare driver going the wrong way; a former partner's threat; drugging symptoms.

Harmless hard cases: banter ("stop it, that's hilarious"); flirting; a horror film retold ("I'll kill you"); a crime report on TV; someone else's story retold; board-game trash talk; a gym trainer ("don't stop"); ordering food; walking home with a friend.

How it was made

• Voices: Kokoro-82M (Apache-2.0), one voice per role; the owner is always af_heart. No real person's voice is used.
• Backgrounds are generated in code (Watch Me's app/scenarios/build.py): party (layered synthetic chatter, a beat and pink noise), bar, street (brown noise with passing cars), TV (a synthetic newsreader across the room) and quiet room tone.
• Scripts were written by the people who built Watch Me. The public Notice cases are paraphrased from official training material; their character names are left out.

Results on this set

Watch Me, streamed through the real app in 4-second chunks (result files):

What hears · what judges	Harmful asked	Harmless asked (needless)
The script word for word · Muse Spark Open (open weights)	11 of 11	0 of 9
The script word for word · hosted Muse Spark 1.3 + 1.1	11 of 11	0 of 9
The script word for word · phrase list and fixed rules, no model	11 of 11	2 of 9
Whisper on the audio · Muse Spark Open	8 of 11 (19 of 22 started)	0 of 9
Whisper on the audio · hosted Muse models	9 of 11 (20 of 22 started)	0 of 9

Notice reviewer on notice_public: Muse Wonder 1.0 40 of 40, Muse Spark Open 38 of 40.

Limits

• Small and synthetic. 22 scenes and 56 cases, written by the system's own builders, with synthetic voices. Use it to compare versions and systems, not as a claim about real-world accuracy.
• English, US-centred. The official sources are US federal and New York State material.
• Rescinded guidance. 11 public cases come from the EEOC's 2024 harassment guidance, which the EEOC rescinded in January 2026; they are used as realistic scenarios with a stated conclusion, not as current law.
• Not for surveillance. The scenes model someone who switched an assistant on for herself. Don't use them to build systems that listen to people without their consent.

Use

from datasets import load_dataset
scenes = load_dataset("hyper07/watch-me-safety-scenes", "scenes", split="train")   # audio needs torchcodec
cases = load_dataset("hyper07/watch-me-safety-scenes", "notice_public", split="train")

Licence and sources

CC BY 4.0. Voices: Kokoro-82M (Apache-2.0). Public cases paraphrased from: U.S. EEOC guidance and Youth@Work material; New York State Sexual Harassment Prevention Model Training (2023) and Case Studies (2018); StopBullying.gov. Each row names its source and links it.