A ~90-minute workshop on working with LLM base models — the raw next-token predictors behind chat models — using the ACS base-model platform. You'll go from "what even is a base model" to reading a model's activations and steering it to talk like a pirate, all over a plain HTTP API. No GPU, no local setup.
- Click the badge — the notebook opens in Colab (nothing to install).
- Get access: at a live workshop, use the invite link your instructor shares — it's instant. Otherwise sign up at infra.acsresearch.org: during the private beta an admin reviews new sign-ups and emails you once you're approved. Either way, copy your API key when it's shown (just once).
- Run the cells top to bottom; fill in the ones marked ✏️ EXERCISE — we do those together.
| File | For |
|---|---|
workshop_student.ipynb |
participants — exercise cells left as TODOs |
workshop_solutions.ipynb |
instructor — every cell filled in |
build.py |
source of truth — regenerates both notebooks (python build.py) |
- Why base models — simulator view, no persona, the distribution as the object of study
- Talking to one — completions, sampling (temperature), logprobs, scoring text with
echo+prompt_logprobs - Reading its mind — pull the residual-stream activations over the API
- Steering it — build a "pirate" direction (difference of means) and add it back mid-generation
- Finale — build a steering vector for a style you choose
- 🎁 Bonus (optional / take-home) — negative steering, a steering-free logprob probe, and open-ended ideas to explore on your own
- Full API + activation docs: https://infra.acsresearch.org/tutorial
- The platform serves open base models behind an OpenAI-compatible
/v1/completions, plus activation harvesting & steering.
Editing: change build.py, run python build.py, commit both .ipynb.