Roadmap¶

The first complete edition covers the full lifecycle from corpus to application. “Complete” here means every stage has an explanation, source trail, and connection to the rest of the system—not that a rapidly changing field has stopped changing.

Maintained surfaces¶

• primary-source and dataset ledgers;
• runnable compact implementations and tests;
• rendered Mermaid diagrams and original visual assets;
• versioned model, dataset, and code maps;
• learning paths and chapter exercises;
• strict documentation build in continuous integration.

High-value future extensions¶

• multilingual tokenization and data-mixture case studies;
• multimodal encoders, projectors, and unified token spaces;
• state-space and hybrid architectures;
• deeper evaluation statistics and contamination labs;
• a small multi-GPU distributed lab;
• reproducible fine-tuning on a redistributable toy instruction set;
• translated editions with terminology review by native speakers.

Edition policy¶

Corrections land continuously. Material architecture or dataset updates should name the applicable version and date. Historical approaches remain when they teach a mechanism still visible in modern systems.