fly.ai compute · training

Train on the fleet

Miners' browsers train small models together. Each generation or round is a batch of jobs that two miners must agree on; the server takes the step in between and keeps every checkpoint. You pay only for jobs that settle.

Mine Buy compute Train Stake Claims Leaderboard Results
START A RUN

What to train

OUR RUNS

What the project trains on the fleet

FROM CODE

Python, and a free practice fleet

pip install "git+https://github.com/alextitonis/fly.ai#subdirectory=mine/python"

from flyai_compute import Fleet, mlp
fleet = Fleet()
run = fleet.train(mlp([784, 128, 10]), x, y, classes=10, eval=(xt, yt), rounds=50, wallet="0xYOU")
print(run.pay_link)
for p in run.watch(): print(p["step"], p["eval_acc"])

Try any run for free first: node scripts/local-fleet.ts in mine/ starts a local server with practice miners. The templates are in the examples (es-cartpole, train-logreg), and the guide has every field.

What fits: up to 100,000 evolved parameters, nets up to 1,000,000 weights, and text from 0.5B to 1.5B models. Miners can see the data you send, so don't send anything private.