Installation

pip install mllm-shap

Optional backends, estimators and visualization:

pip install "mllm-shap[transformers]"   # Hugging Face text backend
pip install "mllm-shap[lmstudio]"       # LM Studio SDK
pip install "mllm-shap[api]"            # OpenAI-compatible HTTP API
pip install "mllm-shap[viz]"            # matplotlib token charts
pip install "mllm-shap[proxy]"          # scikit-learn: the ProxySHAP estimator
pip install "mllm-shap[audio]"          # audio decoding and alignment
pip install "mllm-shap[semantic]"       # sentence, BERTScore and NLI encoders
pip install "mllm-shap[audio-semantic]" # SpeechBERTScore encoder and recognizer

The core package does not install PyTorch. Python 3.12 or newer; uv for development.

Why mllm-shap

• Honest estimands. Shapley and Banzhaf are separate, labelled aggregators; every result and archive records which one produced it, and results report base_value and grand_value so efficiency (sum(values) == v(N) - v(empty)) can be verified; axiom tests enforce it.
• Reproducible runs. One validated configuration, a run id derived from it, and an archive of every model generation, so attributions can be recomputed with another value function without calling the model again.
• Pluggable everything. Partitions, absence policies, estimators, value functions, backends and renderers are entry-point plugins; exact, Monte Carlo, complementary, Neyman, Leverage SHAP, Kernel Banzhaf and ProxySHAP estimators ship built in.

Quickstart

import math

from mllm_shap import (
    ConversationSnapshot,
    ExplainConfig,
    ExplainRunner,
    Message,
    Role,
    Turn,
)

snapshot = ConversationSnapshot.from_turns((
    Turn(messages=(Message(role=Role.USER, text="refund my order"),)),
))
config = ExplainConfig.model_validate({
    "backend": {"kind": "mock", "model_id": "stub"},
    "explanation": {"estimator": "exact", "value_fn": "tfidf_cosine"},
})
result = ExplainRunner(config).explain_sync(snapshot).result

print(result.values)  # one attribution per token
print(result.base_value)  # v(empty): payoff of the fully masked prompt
print(result.grand_value)  # v(N): payoff of the full prompt
# Efficiency: Shapley values sum to v(N) - v(empty).
assert math.isclose(sum(result.values), result.grand_value - result.base_value)

The mock backend needs no model; swap backend.kind for transformers, lmstudio or api to explain a real model. Runs are archived under ./runs/{run_id}; running the same configuration twice raises ArchiveExistsError unless explanation.archive.on_exists is overwrite (or archive.path is "").

Command line

The same run without Python: write a snapshot as JSON, start from the commented config skeleton, and get the result as JSON (--render token_bar --png out.png also saves a chart).

echo '{"turns": [{"messages": [{"role": "user", "text": "refund my order"}]}]}' > snapshot.json
mllm-shap config --example > config.yaml
mllm-shap explain --config config.yaml --snapshot snapshot.json --out result.json
mllm-shap reanalyze --config config.yaml --archive runs/<run_id>   # rescore, no model calls
mllm-shap plugins

See the command-line guide.

Documentation

Full documentation: pawlo77.github.io/MLLM-SHAP

• Getting started
• How a run works
• Configuration
• Choosing an estimator
• Backends
• Extending
• Theory
• API reference

Examples, vault and development

• Runnable notebooks: examples/, also embedded in the docs.
• Architecture notes, decisions, known issues and papers: the Obsidian vault in knowledge/.
• Development setup, checks and pull requests: CONTRIBUTING.md.

Citing

If you use mllm-shap in research, please cite it (CITATION.cff; GitHub's "Cite this repository" button renders it as APA or BibTeX):

cff-version: 1.2.0
title: "mllm-shap: Shapley-value explainability for conversational language models"
type: software
authors:
  - family-names: Pozorski
    given-names: Paweł
repository-code: "https://github.com/Pawlo77/MLLM-SHAP"
license: Apache-2.0

License

Apache License 2.0. See LICENSE.

Copyright 2026 Paweł Pozorski.