Overview

litcoder is a modular Python library for building brain encoding models from naturalistic stimuli (text and speech). It emphasizes clean abstractions and composability:

• Assemblies: load and structure datasets (narratives, LPP, Lebel styles)

• Feature extractors: text and speech encoders with caching and multi-layer support

• Downsampling: align features to fMRI TRs

• FIR expansion: build time-lagged feature spaces

• Models: linear ridge, nested CV, and scikit-learn wrappers

• Training: a dependency-injected AbstractTrainer orchestrates the pipeline

• Logging & plotting: WandB/TensorBoard, brain plots, and run management

Design principles

• Dependency injection for testability and reuse

• Clear data contracts between components

• Deterministic, disk-based caching for expensive features

• Separation of concerns (I/O, processing, modeling, orchestration)