HF Details ML Intern Playbook: 0.8B Distilled Model for $16, All Prompts Open-Sourced
Gradio · x · 2026-10-10
Hugging Face published a detailed writeup of training six models with ML Intern, including prompts, agent runs and full bills.
Key case
- Qwen-Image 2.1's prompt rewriter is a 9B model needing 20GB memory; the author had ML Intern distill it into a 0.8B model that runs on CPU, with 99.7% valid outputs and 1/4 the teacher's token usage. The 9B teacher labeled 8,797 examples; total compute cost: $16.
Workflow & prompting
- Each project starts as one HuggingChat message with ML-intern enabled and ends as a public Hub model with evaluations.
- Agent pipeline: plan → request budget before spending → smoke-test → train → evaluate → publish.
- Prompting style: one-line idea + rationale, then name exact dataset/base model/training script; verified facts go under a heading literally saying "Verified facts, do not re-derive" so the agent spends budget on real work.
- Prompts grew from 450 words (first project) to 2,000 (sixth); all seven prompts are open-sourced on GitHub (yvrjsharma/ml-intern-prompts).
Related event: Hugging Face's ML Intern Turns One Prompt into Trained Models for ~$100(3 posts)→
More from coding & agent
- Google's ambient quality agent diagnoses production AI agents that pass all health checks — rseroter · 2026-10-10
- Running a voice agent from a Tesla screen to book podcast studio slots by phone — altryne · 2026-10-10
- Stanford CS146S Lecture Breaks Down How Always-On Proactive Agents Like Bot Work — soleio · 2026-10-10
- v0 launches Teams: view teammates' chats and build together — iamsahaj_xyz · 2026-10-10
- Same-day Python 3.15 support in uv took deliberate engineering, Astral explains — charliermarsh · 2026-10-10
- Python 3.15 ships, available same-day in uv 0.13.0 — charliermarsh · 2026-10-10