Hugging Face shows how an agent fine-tuned six custom models for $103

A team that needs a small, task-specific model has usually had to pick between an off-the-shelf large one and the half-day of work to train a smaller one. Yuvraj Sharma at Hugging Face published a walkthrough of six models he fine-tuned and shipped in a few days by prompting an agent called ML Intern.

The agent runs inside HuggingChat with ML-intern mode switched on. Hugging Face says it plans the work, asks for a dollar budget before it spends anything, runs a smoke test before the real training job, then trains, evaluates and publishes to the Hub on Hugging Face hardware. The seven prompts behind Sharma's models are open at yvrjsharma/ml-intern-prompts on GitHub.

What the six models cost

Each project landed as a public model with its evaluation in the card, and Hugging Face reports the compute charges for every run:

• Citrus Doctor, a fine-tune of Qwen3.5-2B on 3,017 annotated images of citrus pests and nutrition problems. The base model named the right problem on 14.9% of 335 test photos; after two epochs on one A10G, it got 52.8%. About $1.90.
• Huggy LoRA on FLUX.2 klein base 4B, trained on 84 captioned drawings. About $7.60.
• Pocket Rewriter, a 0.8B student of the 9B prompt rewriter that ships with Qwen-Image 2.1. It returns valid output 99.7% of the time and uses about a quarter of the teacher's tokens, and ships as an 812 MB GGUF file that runs on a CPU. About $16.
• Viewpoint Orbit LoRA, a camera-angle add-on for Qwen-Image 2.1 trained on 1,844 pairs across 23 instructions. About $16.
• Doodle-in LoRA for Qwen-Image 2.1, which replaces a magenta scribble with a named object. Paired with the Viggle turbo LoRA at six steps, it placed the right object in 67.5% of 160 test images at 4.7 seconds per edit. About $24.
• Agate 4-step, a distillation of Logolabs' 260M Agate Preview 002 from 50 steps to 4. The 4-step student scored 0.536 on GenEval against the teacher's 0.563 at 50 steps. About $37 across two runs.

The total for the six models was about $103 in GPU and CPU job charges.

The prompt pattern Sharma recommends

Two lines in the prompt matter more than the rest. The first asks for a baseline before any training, so a reader can see the gain. The second is a 50-step smoke test with a check that the saved weights have actually changed, before any paid run begins. A section labelled "Verified facts, do not re-derive" tells the agent which pieces have already been confirmed, so it spends its budget on the work. Each prompt caps the total spend and asks for permission before exceeding it. The agent begins every task with a zero-dollar budget and refuses to run paid jobs without the cap.

Three things to keep in mind before copying the pattern: the charges are Hugging Face's hardware rates for one developer's runs on A10G and A100 GPUs, not a quote for a team; the models shipped are small enough to fit a specific job but are not replacements for frontier models; and the walkthrough is one developer's account, so the next team's run will look different. Each project also took its share of failed jobs, with the camera-angle LoRA counting 48 jobs across a half-day.

Source

The walkthrough is at Hugging Face, by Yuvraj Sharma and Abubakar Abid. The example prompts are at yvrjsharma/ml-intern-prompts.