← Back
Hugging Face Blog

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

•9 min read•
#agents#llm
Level:Advanced
For:AI Engineers
✦TL;DR

The author used an AI‑agent (ML‑Intern) to distill a 9 B Qwen‑Image 2.1 prompt‑rewriter into a 0.8 B CPU‑friendly model that achieves 99.7 % valid output while using only a quarter of the teacher’s tokens, at a total compute cost of USD 16. By iteratively refining prompts (from ~450 to ~2,000 words) and enforcing budget caps (e.g., USD 12), the workflow produced seven public models, each with baseline zero‑shot scores and smoke‑test checks, including one that more than tripled the accuracy of the Qwen3.5‑2B model. This demonstrates a reproducible, low‑cost pipeline for rapid model specialization using LLM‑driven planning and execution.

⚡ Key Takeaways

  • ML‑Intern orchestrates dataset selection, base model choice, training script, and budget enforcement.
  • Budget caps (e.g., “Cap total spend at USD 12”) prevent runaway costs; total project cost was USD 16 for labeling 8,797 examples.
  • Prompts must include a baseline zero‑shot evaluation line and a smoke‑test step (e.g., “50 training steps” with weight‑change check).
  • Models are published on Hugging Face with evaluation cards; missing “Verified facts” sections still yielded strong results (tripled Qwen3.5‑2B accuracy).

🔧 Tools & Libraries

ML‑InternHugging Face HubClaudeQwen‑Image 2.1Qwen3.5‑2B
💡 Why It Matters

Engineers can now spin up domain‑specific, lightweight models for CPU inference at pennies per experiment, using an LLM‑driven agent to automate data labeling, training, and evaluation while strictly controlling spend.

✅ Practical Steps

  1. Write a detailed prompt (≈2,000 words) that lists dataset, base model, training script, includes “Verified facts” and specifies a baseline zero‑shot metric and a smoke‑test (e.g., 50 steps with weight‑change verification).
  2. Run the prompt through ML‑Intern, approve its budget request (e.g., USD 12 limit), and let it execute labeling, training, and evaluation on Hugging Face hardware.
  3. Publish the resulting model on the Hugging Face Hub with a model card that records the baseline score, training details, and spend.

Want the full story? Read the original article.

Read on Hugging Face Blog ↗

More like this

Censored vs. Truncated Regression in Python, Explained

Towards Data Science•#llm

New in AWS Quick: More capable flows, an updated web experience, and agents that work together

AWS ML Blog•#agents

How Google Duplex evolved to also call businesses for you in Gemini

Google Research•#agents

Q&A: Enabling more human-like interactions and understanding in multimodal foundation models

MIT News AI•#llm

EXPLORE AI NEWS

Daily hand-picked stories on LLMs, RAG, agents and production AI — curated for engineers who ship.

BROWSE NEWS

GET THE WEEKLY DIGEST

Join engineers getting the Monday signal-over-noise AI breakdown. No spam, unsubscribe anytime.

LEARN AI ENGINEERING

Curated courses, research papers, repos and tutorials built for engineers leveling up in AI.

START LEARNING