Hugging Face ML Intern demonstrates a workflow from chat to six custom models

What is changing?

Hugging Face shared results from six custom models produced with ML Intern over several days. The mode inside HuggingChat takes a user's objective, plans the work, requests a budget before spending, runs a small test, and then handles training, evaluation, and publication to the Hub.

The examples span more than one model type. A Qwen3.5-2B adaptation for citrus disease recognition improved from 14.9% to 52.8% across 335 test images and cost about $1.90 to train. A Doodle-in LoRA for Qwen-Image 2.1 placed 67.5% of marked objects in the intended location while reporting 4.7 seconds per edit.

The team also produced a 0.8B student prompt rewriter that runs on a CPU from a 9B teacher, plus a four-step Agate model that runs in the browser. Reported GPU and CPU job costs across the six projects totaled about $103. Model cards and live Hugging Face Spaces demonstrations were published.

Hugging Face recommends a detailed task brief, a pre-training baseline, a smoke test that confirms the weights actually changed, and an explicit cost ceiling. Those controls show that the tool does not remove human oversight; it makes experimentation more repeatable and its budget more visible.