No unlabelled data

tinydistill synth asks the teacher to write example inputs for each label, from the task's own instruction:

tinydistill synth task.yaml --per-label 500 --teacher Qwen/Qwen2.5-7B-Instruct --load-4bit --out texts.jsonl
tinydistill run task.yaml --data texts.jsonl --test test.csv --out run

This is the only place where the teacher generates text, and only inputs: the file has no label field, and the texts are labelled by the usual scoring when you pass them as --data.

Generated texts are much less varied than real ones. On financial tweets, a student trained on about 1,500 generated texts reached 0.596 accuracy, against 0.825 with real tweets (Benchmarks). Use synth to start without data, and replace the texts with real ones as soon as you have some.