ML-intern Workflow Overview
On October 8, 2026, Yuvraj Sharma detailed a workflow utilizing ML-intern, available in HuggingChat, for designing, training, evaluating, and publishing custom models. The agent is structured to plan work, request a budget prior to spending, conduct smoke tests, and subsequently train, evaluate, and publish on Hugging Face hardware.
Prompting and Control Mechanism
Sharma emphasized that the initial prompt is crucial and can be extensive. Example prompts are accessible on GitHub at yvrjsharma/ml-intern-prompts. Prompts may include a section labeled “Verified facts, do not re-derive,” a request for baseline performance (to assess zero-shot capability before training), and a smoke test with an attached verification check. Additionally, deliverables and a defined cost cap are specified; ML-intern initiates tasks with a zero-dollar budget and seeks permission before executing any paid tasks.
Project Highlights and Costs
Sharma reported the development of seven projects within a few days, with notable examples including:
- Pocket Rewriter: A 0.8B student model distilled from a 9B teacher (the official Qwen-Image 2.1 rewriter is a 9B model requiring approximately 20 GB of memory). The 0.8B model operates on a CPU as an 812 MB GGUF file, delivers valid output 99.7% of the time, and utilized about one-quarter of the teacher’s tokens. The total compute cost for labeling 8,797 example requests and training was around USD 16.
- Citrus-disease VLM: This project involved a dataset of 3,017 annotated images across 21 classes used to fine-tune Qwen3.5-2B. The base model achieved a score of 14.9% on 335 test photos; after two epochs on one A10G, the fine-tuned model improved to 52.8%. The compute cost was approximately USD 1.90.
- LoRAs and Image Edits: A Huggy LoRA based on FLUX.2 klein base 4B was trained using 84 drawings at a cost of about USD 7.60; a camera-angle LoRA for Qwen-Image 2.1 utilized 24,722 rendered transparent images and ran approximately 2,000 steps on an A100, costing around USD 16; a Doodle-in LoRA employed 6,042 training pairs at a cost of about USD 24.30.
- Agate 4-step: The distillation of Agate Preview 002 into a 4-step student involved two runs and exports to ONNX for browser compatibility, with combined compute costs across runs reported at about USD 37.
Sharma provided a cost summary table, totaling GPU and CPU charges for all projects at approximately USD 103. He noted that ML-intern conducts small tests before full runs, records licenses and authors for source materials, and publishes model cards and evaluation results upon job completion.
Next Steps for Users
Concrete next steps include enabling ML-intern mode in HuggingChat, pasting a prompt with a budget and baseline request, and consulting the example prompts available on GitHub.
Original source: Hugging Face Blog