Examples¶

The examples/ directory contains Jupyter notebooks with runnable nlp-shap workflows. They complement guides, theory pages, and the API reference.

Setup¶

Install the package:

pip install nlp-shap

From a local clone:

make install

Use the project .venv as the notebook kernel when working from source.

Toy-game walkthrough¶

The notebook below is rendered inline. It compares Shapley and Banzhaf aggregators on a three-player majority game and shows why estimand labelling matters.

Shapley vs Banzhaf on toy cooperative games

Masking walkthrough¶

Partition a conversation into token players, build shared masked views, and compare delete, pad, and neutral absence policies.

Coalition masking and absence policies

Runtime walkthrough¶

Persist coalition rows, deduplicate repeated masks, and schedule async generation with bounded concurrency.

Runtime archive, deduplication, and async scheduling

Exact estimation walkthrough¶

Enumerate every coalition on tiny player sets and delegate attribution to Shapley or Banzhaf estimand plugins.

Exact estimation with estimand plugins

Gallery¶

estimands_toy_game.ipynb: End-to-end walkthrough of the 0.1.1 estimand API: Shapley and Banzhaf aggregators (labels, coalition weights, aggregation), majority vs additive games, ExplainResult labelling, RunManifest wire format, and plugin entry-point resolution. CPU-only; no optional extras.
masking_views.ipynb: End-to-end walkthrough of the 0.1.3 masking API: token partition, DeletePolicy, PadPolicy, and NeutralPolicy (defaults and custom parameters), MaskedSnapshot, MaskCodec, MaskSpace, plugin registry resolution, and validation guards. CPU-only; no optional extras.
runtime_core.ipynb: End-to-end walkthrough of the 0.1.4 runtime API: run archive persistence, coalition dedup keys, hot LRU cache, and async InferenceScheduler with scheduler metrics. CPU-only; no optional extras.
exact_estimation.ipynb: End-to-end walkthrough of the 0.1.5 exact estimation API: ExactEstimator enumeration, estimate_attributions() estimand delegation, budget guard, and plugin resolution. CPU-only; no optional extras.