verbatome

Editorial method · Sources · Provenance

What is measured, handcrafted, and borrowed.

Verbatome separates established mechanisms, teaching simplifications, and original site material. This page makes that boundary inspectable.

Research anchors

The lessons restate these works with original examples; no figure or passage is reproduced.

  1. 01
    N-gram models and generation by samplingShannon (1948)
  2. 02
    Neural language models and learned vectorsBengio, Ducharme, Vincent & Jauvin (2003)
  3. 03
  4. 04
    Regularities in word-vector spacesMikolov, Yih & Zweig (2013)
  5. 05
    Attention and the Transformer architectureVaswani et al. (2017)
  6. 06
    Limits of attention weights as explanationsJain & Wallace (2019)
  7. 07
    RMSNorm normalizationZhang & Sennrich (2019)
  8. 08
    Rotary position embeddings (RoPE)Su et al. (2021)
  9. 09
    Embedding weight tyingPress & Wolf (2017)
  10. 10
    Decoding degeneration and top-pHoltzman et al. (2020)
  11. 11
  12. 12
    SFT and human preferences (InstructGPT)Ouyang et al. (2022)
  13. 13
    Direct Preference OptimizationRafailov et al. (2023)

Teaching corpus and fixtures

The bilingual demonstration corpora were written for Verbatome. They are deliberately small and repetitive so n-gram counts and BPE merges remain legible. “Eat an orange” and “paint it orange,” identifiers such as #901, the 2D coordinates, the Q/K/V matrices, the position vectors, the MLP weights, the training trajectory, and the adaptation lab’s three stages are handcrafted teaching fixtures: they illustrate a computation or a documented behavior and are not measurements extracted from a trained model.

Fonts and imagery

Bricolage Grotesque, Literata, and JetBrains Mono are distributed by Fontsource under the SIL Open Font License 1.1. The SVG logo, social cards, and touch icon are original project assets. The PNG files are generated from the versioned HTML templates in design/og; the site embeds no third-party photography or illustration.

Review policy

Every lesson records a review date in its metadata and states important limitations beside the relevant instrument. A date records a content review, not validation by learners. Moderated learner sessions and manual accessibility review remain separate release gates.

Registry reviewed on 2026-07-18.