Training series
Materials for the training series of the AI for Data – Data for AI program: session plans, slide decks, notebooks, worked-example assets, demonstration scripts, and participant resources. Each session has its own folder, and that folder holds all of the session's materials.
| Session | Title | Format |
|---|---|---|
| 01 | Building AI applications for research and data | 90 minutes, slides and guided notebooks, plus 15 to 30 minutes of questions |
| 02 | Building specialized small AI using synthetic data | Planned |
| 03 | Evaluating AI systems | Planned |
| 04 | Applying AI to metadata curation and data quality | Planned |
| 05 | Building trustworthy AI applications | Planned |
| 06 | All about context: Why AI is really a search problem | Planned |
Session 01 notebooks
They open in Google Colab from the main branch of worldbank/ai4data, and each ends with a link to the next one. The lab (07) needs a T4 GPU runtime, and the notebook explains how to select one.
Conventions
- One worked example per session. Every notebook and every slide of a session uses the same question, the same data, and the same documents, so that what changes between steps is the capability being added.
- Runnable without credentials. Every notebook runs in a fresh Colab environment with no API key. Model calls fall back to recorded outputs that are labelled as such; everything else, from the vector store to the MCP server, runs in full. A live key, when available, is read from the environment and never written into a notebook.
- Popular libraries over custom code. Sessions use the libraries participants will meet in their own work (LangChain, LangGraph, FastMCP, Pydantic, and the program's own packages such as
pcn-core), and write custom code only for what those do not provide. - Real data, labelled. Statistical values come from a named source with the retrieval date recorded. Documents written for a workshop say so in their header, state where their figures come from, and are never presented as publications.
- Cached outputs. Notebooks are committed with their outputs so that reviewers and participants can read them without running them. The regeneration script reruns them in fallback mode.
- No names in the materials. Authors and reviewers are recorded in the session README under roles until the material is released.
- No confidential documents. The corpus of a session is public or written for the session.
- House style. The writing rules of the cookbook authoring guide apply to slides, notebooks, and resources: descriptive titles, plain statements, no slogans, no antithesis as a habit.
Source files
The pages of this section are generated from the training folder of the repository. The slide decks, the demonstration scripts, and the scripts that regenerate the notebooks are kept there.