Argilla’s new tool lets you build AI training data by just describing it

Argilla just shipped a synthetic data generator that turns natural language descriptions into ready-to-train datasets. If you’ve ever stared at a blank spreadsheet dreading the slog of manual labeling, this is aimed squarely at you. The tool uses a pipeline powered by distilabel and the free Hugging Face text-generation API, but the whole point is that you never have to touch that backend. You describe what you want — say, a customer-support chat dataset for a fintech app — and it spits out samples you can refine through a three-step UI.

The generator currently handles text classification and chat datasets, the two formats that feed most fine-tuning workflows. Classification covers the classic use case: categorizing reviews, social posts, or news articles. Argilla’s own demo dataset, argilla/synthetic-text-classification-news, sorts synthetic news into eight categories. Chat datasets target supervised fine-tuning, the process that makes LLMs conversational. Their customer-support example is a single-turn dataset where the support topic is, fittingly, the synthetic data generator itself. Throughput sits at about 50 samples per minute for classification and 20 for chat on the free API — decent for prototyping, though anyone scaling up will want to plug in their own Hugging Face account and custom models.

What makes this more than a toy is the integration with Argilla’s core collaboration platform. Once you generate a dataset, you open it directly in Argilla to review, filter, and curate it with semantic search before exporting to the Hugging Face Hub. That review step is where the real work happens, because synthetic data is never clean on the first pass. The generator produces a sample, you tweak the system prompt and task settings, regenerate, and iterate until the outputs match your standards. From there, Argilla nudges you toward AutoTrain for a no-code fine-tuning run — pick a task, point it at your new dataset, and you’ve closed the loop from idea to deployed model without writing a line of Python.

I’ll be watching whether the promised RAG and evaluation dataset support actually lands. Those are harder problems; generating realistic retrieval-augmented generation examples or evaluation benchmarks demands more structural rigor than freeform chat or classification. But for the core use case — quickly bootstrapping a labeled dataset when you have zero examples — this lowers the barrier enough that a product manager or domain expert could run the whole pipeline. That’s genuinely useful, even if the resulting model will still need human vetting before anything customer-facing.