Cookbook
Each recipe runs a whole workflow with your own key, while an example shows one feature in a few lines. Every recipe except the workshop notebook is a file that you download. You open the workshop notebook in Colab.
| Recipe | What it does | What you need | Run it |
|---|---|---|---|
| Steer a model with a contrast vector | Builds a steering vector from captured activations that makes llama-8b write like a pirate, and measures its effect at each scale |
Python 3.11 or later, uv |
uvx marimo edit --sandbox --watch pirate_steering.py |
| Workshop notebook | Takes you from a first completion to a steered model, with exercises on sampling, logprobs, and activations | A Google Account and an API key | Open in Colab |
| Submit and poll a completion job | Gets a completion that keeps running on the server after a dropped connection, and returns the same job when you rerun it | Python 3.11 or later, httpx |
python python_completion_jobs.py --model llama-8b --idempotency-key ... |
| Stream through a cold boot | Streams a completion from a cold model and keeps one request open until the first token arrives | Python 3.10 or later, uv |
uv run python_long_requests.py --model llama-405b |
The downloaded files read your key from ACS_API_KEY and the API address from ACS_API_BASE. Without ACS_API_BASE, they call the production API. In the workshop notebook, the first code cell asks you to paste the key.
Our CI checks the completion-jobs script and the long-requests script against a local test server on each pull request, so they stay runnable. The pirate notebook is not part of that check.
Next steps
- Get a key and send your first request.
- Read logprobs at each generated token for the smallest working example.
- Capture the residual stream before you write your own steering recipe.