What the work involves
You receive Cantonese model outputs — sometimes written, sometimes transcribed or synthesised speech — and judge them against project rubrics for linguistic accuracy, fluency, and contextual appropriateness. Much of the real work is separating genuine errors from stylistic variation: an output written in Standard Written Chinese when the prompt called for colloquial Cantonese, a Mandarin-calqued construction that is grammatical on paper but nobody would say, a particle cluster in the wrong order, a Jyutping romanisation that doesn't match the tone. You then write the rationale, because a rating with no explanation is unusable to the team training the model.
Alongside evaluation, expect annotation, labelling, and categorisation batches shaped to NLP needs, plus ad hoc quality reviews when priorities shift. You'll raise edge cases with project leads — regional variation between Hong Kong, Macau, and Guangzhou usage, or disagreements over where the line between "awkward" and "wrong" sits — and help tighten the guidelines rather than quietly inventing your own.
What the screen looks for
micro1 runs an AI-led interview. It probes whether your Cantonese is genuinely native across registers, not conversational-only: whether you can articulate written Cantonese versus Standard Written Chinese, name concrete interference patterns from Mandarin or English, and stay consistent when a follow-up pushes back on your judgement. No AI background is required, but vague answers about "sounding natural" get filtered out fast. Being able to point to prior translation, subtitling, transcription, annotation, or teaching work helps.
Logistics
- Fully remote contractor engagement, asynchronous task queues
- Volume varies with client demand; treat it as flexible part-time rather than guaranteed hours
- Requires your own computer, stable internet, and usually a decent headset for audio tasks
- Observed band is $30–40/hr; rate depends on project and task type and is not guaranteed