coreweave/cwsandbox-recipes

coreweave/cwsandbox-recipes

Description: Examples for running AI workloads in CoreWeave Sandboxes

Language: Python

License: Apache-2.0

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-09-11T21:26:07Z

Pushed: 2026-09-11T21:36:48Z

Default branch: main

Fork: no

Archived: no

README:

CoreWeave Sandbox Recipes

Practical, runnable examples of AI workloads on CoreWeave Sandboxes.

Sandboxes are isolated compute environments with serverless and CoreWeave Kubernetes Service (CKS) placement options. Create CPU or GPU workloads from Python with the cwsandbox SDK, not the retired wandb[sandbox] wrapper. See the get started guide.

Recipes

| Recipe | What it shows | Hardware | |---|---|---| | [taubench-verl](recipes/taubench-verl/) | RL training (GRPO) of a tool-using agent on tau-bench with veRL. GPU sandbox or optional SkyPilot/CKS trainer; per-worker CPU sandbox pools, serverless by default. W&B Models metrics, optional Weave traces, and checkpoint reference artifacts. | 8x H100 + CPU pools |

Getting started

1. Pick a recipe and open its README. 2. Copy its .env.example to .env and fill it in. Every variable is documented inline. 3. Follow the recipe's run steps.

Each recipe is self-contained: its own dependency declarations, config, and tests. Follow its installation steps and configure credentials and placement before running it.

Cost and safety

Recipes provision billable cloud resources and may call paid APIs. Read a recipe's resource requests and cleanup section before running it; pool sizes may be per worker, not global. Use an existing bucket for checkpoint uploads and verify persistence before teardown. Keep credentials in .env (gitignored) or a secret store, never in source files.

Contributing

See [CONTRIBUTING.md](CONTRIBUTING.md). Start a new recipe from [_template/](_template/).

License

Apache 2.0. See [LICENSE](LICENSE).