Hugging Face's ML-intern agent trains small custom models only after the user approves a budget

The cap is a sentence in the prompt. One of the author's reads: "Cap total spend at USD 12 and ask me before exceeding it." ^[3] A line of text cannot stop a GPU job by itself. According to the post, the limit holds because the agent starts each task with a zero-dollar budget and needs permission before it executes any paid job ^[2]. Leave the cap out and it proposes a couple of paths sized to the project, then asks which one the user wants ^[4].

The two prices both write-ups repeat sit at the cheap end. Subtract them from the six-project total reported by dev.to and the other four runs come to $85.10, or about $21 each ^[23]. At least one of those four cost more than the $16 rewriter ^[23]. The $1.90 is a compute figure ^[5]. Its 3,017-image dataset was assembled beforehand by the author, with Claude, from three Project-AgML sources ^[6]. The rewriter's $16 covers making its own data: inside that budget, the 9B teacher labelled 8,797 example requests ^[8].

What stayed expensive is the brief. "The first message is where I spend my effort," wrote the post's author, a Hugging Face team member ^[11][18]. Briefs name the dataset, the base model and the training script, and put anything already checked under a heading that says "Verified facts, do not re-derive" ^[20]. For the camera-angle LoRA, that section recorded which trainer had just added transparent-image support and which open GitHub issues made the fallback trainer risky ^[21]. Knowing either is an ML engineer's job. The intern wants a long ticket: the first prompt ran about 450 words and the sixth about 2,000 ^[10].

A thinner brief still worked. The citrus prompt had no verified-facts section, and the fine-tune still beat its Qwen3.5-2B base by about 3.5 times on accuracy ^[14][24]. That comparison exists because the prompt asked for it: "Also report the base model's zero-shot score on the same metric before training so we can see the gain." ^[12] Against its own baseline the gain is real. The same model is wrong on about 47% of its test photos ^[25].

A human still has to decide when a model is good enough, and the agent's evaluation output helps. On a Huggy LoRA trained on FLUX.2 klein base 4B from 84 captioned drawings, it saved a checkpoint every 100 steps and drew the same prompts with each ^[22]. Step 200 was the first fully on-model, and from step 500 the style began bleeding into other prompts ^[22].

In our view the budget gate is the part other agent builders should copy, because a reviewer sees a price before any hardware is billed ^[2]. The brief that keeps the price small still takes someone who knows the training stack. ML-intern is live in HuggingChat behind a mode switch ^[17], and the author's seven prompts for the six models are on GitHub at yvrjsharma/ml-intern-prompts ^[19].