Skip to content

FAQ

First-run questions about SFT versus RL, endpoint selection, and API key access.

Should I use SFT or RL?

Use SFT when you already know the target output and have labeled examples to learn from.

Use RL when there is no single fixed answer but you can score behavior with a reward function, a verifier, a test suite, or environment feedback.

If you have both, you can combine them. A common pattern is to start with SFT for the base behavior, then use RL to optimize further — but this order is not mandatory for every task.

Does Mint Recursive support SFT?

Yes, Mint Recursive supports SFT directly.

The standard SFT path is:

  • forward_backward(..., loss_fn="cross_entropy")
  • optim_step(...)

Which domain should I use?

Choose by your network path:

  • Mainland China → https://mintcn.macaron.xin/
  • Outside Mainland China → https://mint.macaron.im/

If you are unsure, try the domain that matches your region first. Lower latency and a stable connection are what matter in practice.

Where do I get MINT_API_KEY?

Sign up at mint.macaron.im or mintcn.macaron.xin — registration is self-service.