📝 Additional Notes

🧠 System and Prompt Tokens

System messages include both explicit system turns and internal prompts added during model setup.

⚙️ Memory and Performance

During explanation, package stores combined embeddings (after normalization) and mask data used for SHAP sampling. Most intermediate data is dropped when verbose=False. Even then, long conversations and large models can require significant memory.

For larger workloads, prefer approximation strategies (for example Monte Carlo variants with controlled sample budgets).

🔗 External Embeddings

To compare behaviors across models, use external embedding implementations from mllm_shap.shap.embeddings. Embedding calls run sequentially after each generation step (no batching by default), so embedding model throughput directly impacts total runtime.

Warning

Some features are experimental and may change in future releases. Always check the examples/ folder for updated usage patterns.