Examples¶
Everything here is built on toolbox without changing it. Each folder is self-contained: code, configs, tests and a README that says how to run it and what it shows. Its README is also a page on the docs site.
| Example | Area | Source | Folder |
|---|---|---|---|
| Recipes: RLHF pipeline, self-improvement, agent tracing, custom losses and layers | Workflows | toolbox | recipes/ |
| Graph model views (GMV): serve GNN outputs as views via L-bisimulation | Graph inference serving | Fan, Lu, Che & Wu, Approximating Graph Inference by Serving Graph Model Outputs as Views (2025) | gmv/ |
| Jev decision models: distill debiased or reasoned decisions into one forward pass | LLM decision models | Nokia Applied Research, AnyJev (used as a library) | jev/ |
Run an example from the repository root with python -m examples.<folder>.<script>, for
example python -m examples.jev.run. Run its tests with pytest examples/<folder>.
Adding an example¶
Create examples/<name>/ with README.md, run.py, configs/ and tests/, and add a row
to the table above. The README becomes a docs page automatically. For a paper, the README
covers the method, the citation, results, and what is and isn't reproduced;
gmv/ is a good template.