The model that didn't exist, so you made it yourself

Last week, the author wanted a smaller version of the prompt rewriter included with Qwen-Image 2.1. Although official models are 9B in size and require substantial memory, the author found only compressed copies on the Hub. Determined to create a version that runs on a CPU, they instructed the ML Intern to generate a 0.8B model, which uses only a quarter of the resources of the original.

The entire project cost USD 16, including training 8,797 example requests for the 9B model. Over subsequent days, the author created five more custom models through this process. Each model started as a message in HuggingChat with ML-intern enabled, and ended as a public model on the Hub, complete with evaluation in the model card.

ML-intern manages the planning, budgeting, testing, training, evaluation, and publishing of the models. To ensure efficient use of resources, ML-intern begins each task with a zero-dollar budget and requires permission before executing any paid jobs. The author provided detailed prompts for each project, including verified facts, dataset specifications, and expected deliverables.

By leaving out certain sections, ML-intern still produced impressive results, demonstrating the tool's versatility and effectiveness.