Language model labs

Three related kits for working with language models without treating a hosted API as the only option. Start with how apps call models, then train a tiny model locally, then ground a profession twin in editable canon.

How they relate

  1. LLM 101 – consume a model in an app (tokens, tools, RAG, evals).
  2. SLM 101 – train and run a tiny model on your machine (corpus, checkpoint, generate).
  3. Digital Sibling – keep judgment in editable canon and retrieve it; optional local generation later.

For day-one job ask (manuals, unfamiliar languages), prefer Digital Sibling packs plus a local base model. Use SLM 101 to understand training; use LLM 101 to understand the app loop around a model.

LLM 101

Concept lab for how LLM applications work: tokens, fake inference, tool calling, retrieval, and evals. Offline-first exercises; no API key needed for the core path.

SLM 101

Concept lab for training a tiny language model locally: corpus, checkpoint, generate, and eval. Offline-first Python; bootable with Docker Compose.

Digital Sibling

A local-first digital twin of a person and their working life: editable canon, retrieval, intake questionnaire, and a fictional sample persona. No API key for the core path.