Hugging Face introduces ML Intern, an agentic tool for custom model fine-tuning and deployment

Hugging Face has introduced ML Intern, an agentic tool designed to automate the creation of custom models by fine-tuning existing models based on user-defined prompts. The tool handles dataset curation, fine-tuning, evaluation, and publishing on Hugging Face hardware, with built-in budget controls and verification steps. Users describe the desired model in natural language, and ML Intern plans the work, requests budget approval before spending, runs small tests, trains the model, evaluates performance, and publishes the result as a public model on the Hugging Face Hub with an accompanying model card.

The author of the announcement used ML Intern to create seven custom models over a few days, starting with a prompt of about 450 words and refining it to around 2,000 words by the sixth project. Each prompt included verified facts to avoid redundant work, baseline performance requirements, smoke tests with checks, and strict spending limits. For example, the author capped total spending at USD 12 per project and required ML Intern to request permission before exceeding the budget. If no budget is specified, the agent suggests options based on project size and asks the user to choose.

One project involved fine-tuning Qwen3.5-2B to create a citrus disease classifier. The author assembled a dataset of 3,017 annotated images across 21 pests, illnesses, and treatments from the Project-AgML organization on the Hub. ML Intern fine-tuned the model on one A10G GPU for two epochs, increasing accuracy from 14.9% to 52.8% on a 335-image test set at a cost of about USD 1.90. The resulting model, citrus-disease-vlm-instruct, and its dataset are publicly available, along with a demo app called Citrus Doctor.

Another project used ML Intern to create a character-specific LoRA for the FLUX.2 klein base 4B model. The author trained on 84 captioned drawings from the Chunte/huggy_for_training dataset. ML Intern saved checkpoints every 100 steps and generated images with each checkpoint, enabling easy selection of the best model. The LoRA worked on the distilled klein model at 4 steps. The total compute cost was about USD 7.60. The model, dataset, and a demo app called Huggy Generator are publicly available.

The author also used ML Intern to create a camera-angle LoRA for Qwen-Image 2.1, a capability not yet available in the community. ML Intern rendered 1,030 scanned household objects from Google Scanned Objects at 24 angles each, producing 24,722 transparent images on a CPU job that cost a few cents. It then finalized 461 objects for training and 40 for testing, with 1,844 before-and-after training pairs. Training ran 2,000 steps in about 90 minutes on one A100 GPU at a cost of about USD 3.75, for a total project cost of about USD 16. The model, dataset, and a demo app called Viewpoint Orbit are publicly available.

In a separate project, the author tasked ML Intern with creating a doodle-in LoRA that replaces a magenta scribble in a photo with a named object while preserving lighting and composition. No dataset existed for this task, so the author described how to create one in the prompt. ML Intern wrote and tested pair-building scripts in a CPU sandbox, then ran them as GPU jobs, recording the source and license of every image. It built 6,042 training pairs and a 160-pair test set, with 40 pairs from 23 unseen object classes. Training ran 2,000 steps in 1 hour 38 minutes on one A100 GPU at a cost of about USD 4, for a total project cost of about USD 24. The model and dataset are publicly available.