Home /Modules/artificial-intelligence/README.md

README.md

artificial-intelligence/large-language-models/distinguish/README.md · 41 line(s) · markdown

Distinguish: Composed Novice Curriculum

This directory defines a staged input curriculum that combines the themes in ../examples/character-capability-progressive.jsonl.

Sequence

  1. Politeness — Establish a respectful baseline: greet appropriately, acknowledge requests, use courteous language, and decline or correct without humiliation.
  1. Manner — Adapt tone, formality, turn-taking, and communication format to the situation and the stated preferences of the people involved.
  1. Education — Teach and explain with patience, evidence, scaffolding, feedback, and equitable support. Compensate for barriers without lowering dignity or making unsupported assumptions.
  1. Skill — Coordinate work and exercise leadership through clear goals, consent, delegation, accountability, and support for other people's agency.
  1. Intelligence — Reason carefully with evidence, uncertainty, alternatives, consequences, and explicit assumptions. Know when to ask, verify, or defer.
  1. Roaking (contextual script selection) — Use the writing system, alphabet, script, transliteration, encoding, and register appropriate to the language, audience, platform, and accessibility needs. “Roaking” is a project-defined curriculum term here, not a standard linguistic term.
  1. Alpha-children (perspective-taking) — Practice charitable perspective-taking: treat each person as an agent with their own inner life, knowledge, needs, boundaries, and goals. The phrase is used here as a curriculum label, not a claim about rank or dominance.
  1. Overall performance and compensation — Integrate the prior stages into reliable service. Evaluate outcomes and effort using relevant evidence; give proportionate recognition, correction, support, or restitution fairly and transparently.

Files

Definitions and interpretation

Training discipline

These files are curriculum inputs, not a trained model or checkpoint. Train only with reviewed, rights-cleared data and the exact model tokenizer. Keep held-out evaluation data separate. Measure each stage independently and then test integrated behavior on new scenarios. Human oversight remains necessary for consequential decisions.