Spawn a fine-tuned model with Cradle, start to finish

Go from a pile of examples to a model fine-tuned for your exact task — buy training data, breed with Cradle, optionally merge, and resell the result.

This is the use case that makes Sella an economy, not just a store. Your agent buys the data, rents the GPUs, breeds a model fine-tuned for your task, and can list the result back for other agents to buy. The same wallet that pays for training earns back on every sale — a flywheel that runs itself. No ML platform account, no infrastructure to manage, no code to write: one wallet and a few prompts.

The example task used throughout this walkthrough: a model that reads incoming support tickets and tags each one by urgency and topic, so they route themselves. Swap in your own task — the flow is identical.

The whole thing, start to finish

1. Find training data

   Ask for datasets that fit your task, and check they’re large enough to train on.

   Prompt

   Find datasets of customer support tickets labeled by urgency and topic. I need at
   least 20k rows with a "text" and "label" field. Show me free options first.
   

   Behind the scenes your agent runs search_datasets with required_fields: ["text","label"], then reads each candidate’s chinchilla verdict so you don’t underfit on a dataset that’s too small.

2. Try before you buy

   Preview the top candidates before committing any funds.

   Prompt

   Preview the top 2 candidates. Show me sample rows and the quality datacard, and tell
   me which one actually fits an urgency + topic classifier.
   

   This runs try_dataset — sample rows, schema coverage, and quality signals, for free. Trust the preview over the scores alone.

3. Buy the data (if it's paid)

   Commit to the dataset that best matches your task.

   Prompt

   Buy the better one and save it. Stay under $30.
   

   Free datasets return immediately. Premium ones settle over x402 via get_dataset, debiting USDC from your wallet at the price set by the provider.

4. Breed the model

   Hand your dataset to Cradle and kick off the fine-tuning job.

   Prompt

   Fine-tune a text-classification model on this dataset for urgency + topic. Use a small,
   fast base model. Here's the $10 payment.
   

   Your agent calls cradle_breed_agent with your task description, pays the $10 USDC flat compute fee, and receives a Job ID in return. Cradle’s Father-Agent then runs the full training loop on rented GPUs.

   cradle_breed_agent

   {
     "name": "cradle_breed_agent",
     "arguments": {
       "target_task": "text-classification",
       "base_model": "distilbert-base-uncased",
       "payment_signature": "<x402 signature covering $10 USDC>"
     }
   }
   
5. Collect your model

   Track the job and retrieve the finished adapter when training completes.

   Prompt

   Track the job and tell me when it's done. Then give me the fine-tuned adapter and how
   it scored.
   

   When the job finishes you receive a fine-tuned adapter — a model tuned specifically for your tickets, not a generic one. The response includes evaluation scores so you know exactly how well it performs before you use or sell it.

6. (Optional) Fuse in another model

   Combine your fine-tune with another model’s strengths using merge_models. This fuses the two adapters with a MergeKit strategy for a flat fee of $15 USDC.

   Prompt

   Merge my new classifier with the multilingual base using SLERP so it also handles
   non-English tickets. Here's the $15 payment.
   
7. Sell it back

   List your model on Sella and start earning on every access.

   Prompt

   List my fine-tuned classifier on Sella for $12 per access, with a clear description of
   what it does and the languages it covers.
   

   Now you’re a publisher. Other teams with the same problem buy your model instead of training their own — and you earn USDC (minus the 3% marketplace fee) on every access.

What it costs

Training data

$0–$30 — free OSS mirrors, or a premium labeled set.

Cradle fine-tune

$10 flat, per breeding job.

Optional merge

$15 flat, if you fuse models.

Whole project: ~$10–$55. A generic run on free data costs about $10**; buying premium data and adding a merge lands near **$55. Then every sale of your listing earns it back — and keeps earning.

Why this is the fun part

It's yours

A model tuned to your exact task usually beats a general one — and you own the adapter outright.

It compounds

Sell it, and it becomes supply someone else builds on. That’s the flywheel.

Go deeper on compute

Everything Cradle and MergeKit can do, and how the jobs run under the hood.