Product release ·
Training Lab, now over x402
Training Lab is open to everyone. Teach a small open model your task, pay once, and get it back with a score report. Now an agent can do the whole thing with nothing but a wallet, and developers can do it from code with an API key.
How it works
Your data in. Your model out.
Bring a dataset of example conversations, or describe the behaviour you want and generate one. Pick a base model and a GPU block, and the price is fixed before you pay.
We rent the GPU, train, keep an evaluation split so the score is measured, and shut the GPU down when the job is done.
- 01
Data
Upload a JSONL file of chat examples, or generate one. It is checked line by line before it can train.
- 02
Model and GPU
Qwen 2.5 0.5B or 1.5B Instruct, on the GPU and block length you choose. One fixed price.
- 03
Results
A LoRA adapter and a score report, plus a GGUF when the conversion succeeds. Download them, or deploy as a private AI Machine.
Over x402
An agent can train a model on its own.
No account, no API key, no prepaid balance. The agent asks for a run, gets an HTTP 402 with the exact price, pays in USDC on Base or Solana, or USDG on Robinhood Chain, and follows the run with a claim token.
The request is checked before the 402 and again before the payment settles. If the GPU still cannot start, the payment becomes credits on your account.
- 1 · Catalog
GET /training/x402/catalog- 2 · Quote
- A wallet-signed
POST /training/x402/runsreturns a 402 with the exact price - 3 · Pay
- The same body plus the
quote_id, withX-Payment, returns a claim token - 4 · Follow
GET /training/x402/orders/:idwith the claim token
The paying wallet must own an approved dataset. Agents can buy one over x402 too, or upload one with a wallet signature.
Your money
Built so a payment can only do one thing.
- The price is the quote
- You sign exactly the amount you were quoted, and that is what the run costs. Anything else is refused before money moves.
- One payment, at most one run
- Sending the same payment twice never charges twice. It returns the same order. If the GPU cannot start, the payment becomes credits.
- The right wallet only
- The wallet that pays must own the dataset. A payment from anyone else is refused before it settles.
- Unclear outcome
- If we cannot confirm a payment yet, you are told to wait, never to pay again.
- If the GPU cannot start
- The payment becomes credits on the paying wallet's account, and the status says so.
A run is a prepaid block: finishing early or cancelling does not refund the rest. Up to $500 per order and 3 paid runs per wallet per day over x402.
API keys
From code, with a key that cannot surprise you.
Everything the dashboard does is in the API: upload and approve datasets, estimate a run, start it, follow it, download the result.
Training runs rent GPUs, so they now have their own permission, training:write. New keys made only for datasets cannot start a GPU unless you say so. Keys you already have keep working as before.
Proof
xMESH was trained here.
Singularity xMESH 0.8B, our x402 decision model, was trained and evaluated in Training Lab, and it gets 80% of complete x402 flows exactly right.
Get started
Train your first model.
Open Training Lab in Singularity Cloud, follow the API guide, or give your agent the x402-compute skill and let it handle the rest.
