Where a run lives
Your code. Our GPUs.
The training data stays in your script. The weights stay on the GPUs until you download them.
In your script
You hold thisThe loop runs on your machine. A laptop is enough. The script sends the batch. It does not hold W.
- 01
Prompts and tasks
The data, the traces, and the sandbox those tasks run in.
- 02
Rewards
An exact match, a unit test, a rubric, or a judge you call.
- 03
The loss
Supervised, reinforcement, preference, or a loss you write.
On the GPUs
Two copies of WSmolGym loads the model and keeps both copies. The sampler reads one. The trainer writes the other.
- 01
Sampler
Reads the last synced weights. Does not train.
- 02
Trainer
Writes W when you call
optim_step(). That is the only place W changes. - 03
Sync
Copies the written weights onto the sampler, so the next
sample()sees them. - 04
save_state()The checkpoint. Optimizer state stays, and the weights are yours to download.
On this website
No accountThe site has no sign-in and no analytics, and it does not set cookies. The only personal data is an email you choose to send to swapnil@rycerz.es. That message is the request for a session.
What happens to that email is on the privacy page.