> ## Documentation Index
> Fetch the complete documentation index at: https://docs.x402layer.cc/llms.txt
> Use this file to discover all available pages before exploring further.

# Training Lab

> Teach a small open model your task. Pick your data, pick a model, pay once, and get the model back with a score report. From the dashboard, with an API key, or over x402 with nothing but a wallet.

Training Lab fine-tunes a small open model on your own examples and gives it
back to you: a LoRA adapter, a score report that shows how much it learned, and
a downloadable GGUF when the conversion succeeds (`gguf_available`). You choose the data, the base model and a GPU block. We
rent the GPU, run the job, and shut it down when it finishes.

Singularity xMESH 0.8B was trained and evaluated in Training Lab.

## Three ways in

> **[Dashboard](https://cloud.x402compute.cc/network/training)**
>
> Pick a dataset and a model in Singularity Cloud, review the price, start.

> **[API key](/cloud/training/api.md)**
>
> Upload datasets and start runs from code with an `X-API-Key`.

> **[x402 (agents)](/cloud/training/agents.md)**
>
> No account: get a 402 with the exact price, pay in stablecoin, poll.

## How a run works

1. **Data.** Upload a JSONL file of chat examples, or describe the behaviour you
   want and have Training Lab generate a dataset. A dataset must be approved
   (schema-checked) before it can be trained on.
2. **Model.** Choose a base model from the catalog (today Qwen 2.5 0.5B Instruct
   and Qwen 2.5 1.5B Instruct).
3. **GPU block.** Choose a GPU and a number of hours. The price is fixed before
   you pay: hours x the GPU's hourly training price.
4. **Training.** The run reports progress while it trains. Every run keeps an
   evaluation split so the score report is measured, not guessed.
5. **Results.** Download the adapter (and the GGUF when available), or deploy the model as a private
   AI Machine.

## Pricing and refunds

You pay for a **prepaid block**: the GPU hours you chose, at the hourly price in
the catalog. Finishing early or cancelling does not refund the rest of the
block.

If something on our side stops a run before it trains (the GPU cannot start,
the provider fails, the run never reports in), the charge is refunded to your
credits automatically.

Payments made over x402 are refunded the same way: as credits on the paying
wallet's account, never on-chain.

## Limits

| Limit | Value |
| - | - |
| Longest block | 48 hours |
| Base models | Qwen 2.5 0.5B Instruct, Qwen 2.5 1.5B Instruct |
| x402 order size | Up to \$500 per run |
| x402 runs per wallet | 3 per 24 hours |
| x402 networks | Base (USDC), Solana (USDC) and Robinhood Chain (USDG). `GET /training/x402/catalog` lists what is live. |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.