Use an AI agent (Claude Code) with Vast.ai GPUs and Hugging Face to build models

Training or fine-tuning a model is mostly plumbing: finding a base model, cleaning a dataset, picking a GPU, fixing CUDA and dependency errors, babysitting a run for hours, then evaluating and uploading. That's exactly the kind of long, fiddly, tool-heavy work an agent like Claude Code is good at — as long as it has a real machine and the right accounts connected.

The setup in one picture

The key idea: don't keep a GPU running for the agent to think on. The agent's home is a normal server; GPUs are rented for the minutes or hours of actual training.

What the agent does, step by step

1. Understand the goal. "Fine-tune a small open model to classify our support tickets into 12 categories" or "Make a model that writes product descriptions in our tone."
2. Prepare the data. It cleans your CSVs or exports, removes duplicates, splits train/test, formats examples, and can push the dataset privately to the Hugging Face Hub.
3. Pick a base model. It searches the Hub for suitable open models, checks licences and sizes, and recommends one that fits your budget.
4. Write the training script. Typically Hugging Face transformers with peft (LoRA) or trl for supervised fine-tuning — tested first on a tiny sample on the CPU server.
5. Rent a GPU on Vast.ai. Using the vastai CLI, it searches offers by GPU type, VRAM, price and reliability, starts an instance from a PyTorch image, and copies the code over.
6. Run and watch. It launches training, tails the logs, and fixes the usual problems — out-of-memory errors, wrong batch sizes, missing packages — without you waking up at 3 a.m.
7. Evaluate. It scores the model on the held-out set and compares against the base model, so you see whether training actually helped.
8. Publish and clean up. It pushes the model or LoRA adapter to Hugging Face (private or public), writes a model card, optionally builds a small demo — then destroys the Vast.ai instance so billing stops.

Connecting the accounts

• Hugging Face: create an access token (read for downloading, write for publishing) and log in on the server with the Hugging Face CLI. Use a fine-grained token limited to the repos the agent should touch.
• Vast.ai: add credit, create an API key and install the vastai CLI. Keeping only a modest balance on the account is a simple, hard spending cap.
• Claude: your own Claude subscription or Anthropic API key powers the agent.

Store keys as environment variables or in the CLI's config on the server — never paste them into prompts or commit them to a repo.

Where Google Colab fits

Google Colab is great for poking at a dataset or trying an idea in a notebook with a free or cheap GPU. The agent can write notebooks you open in Colab to explore data and look at results, while the real training runs happen on Vast.ai where runs can last for hours and you choose the exact GPU. Many people use both: Colab to explore, Vast.ai to train.

What does it cost?

Vast.ai prices are set by the hosts and change constantly, but consumer cards like an RTX 4090 are often well under a dollar an hour and data-centre GPUs like A100s and H100s cost a few dollars an hour. A LoRA fine-tune of a small open model on a few thousand examples often finishes in well under a few hours — so the GPU bill for an experiment can be a few dollars. Always check current prices on Vast.ai.

The biggest cost risk isn't the training — it's a GPU left running by accident. Tell the agent, in writing, to destroy instances when a run finishes or fails, and to check for running instances at the end of every session.

Good first projects

• Text classification — tickets, reviews, leads, emails sorted into your categories.
• Extraction — turn messy documents or pages into clean structured fields.
• Tone and style — a small model that writes in your brand's voice for cheap, high-volume jobs.
• Embeddings and search — fine-tune an embedding model so search over your own content gets better.
• Image models — LoRA adapters for a product, a style or a character.

Start small. A fine-tuned small model that does one job well is usually cheaper and faster than prompting a giant model for the same task — and you own it.

Guardrails that matter

• Budget rules: max price per hour, max hours per run, always destroy instances when done.
• Data rules: keep datasets private on the Hub unless you decide otherwise, and don't upload personal data you don't have a right to use.
• Licences: check the base model's and dataset's licence before training for commercial use.
• Checkpoints: push intermediate checkpoints so a lost or interrupted instance doesn't cost the whole run — cheap marketplace GPUs can occasionally go offline.

Why the agent needs its own computer

Training runs are long. A laptop that sleeps, a chat window that times out, or a sandbox that resets will kill the job halfway. An always-on server with a terminal and browser lets Claude Code work for hours, keep logs and datasets between sessions, open the Hugging Face and Vast.ai dashboards when it needs to check something, and report back when you look at your phone.

Common questions

Claude and Claude Code are products of Anthropic. Vast.ai and Hugging Face are independent companies. WebHand is not affiliated with any of them.