v0.1.0 — first public release

v0.1.0 — first public (alpha) release

Analyze and visualize how an LLM picks each next token: probabilities,
attention, logit-lens, activations, manifold geometry, and per-head direct
logit attribution — CLI + static web viewer + native desktop app.

Install

• pip install llm-token-heatmap → the token-heatmap CLI & library
• Or download a desktop installer below (macOS / Windows / Linux)

Highlights

• CLI: trace, diff, manifold, serve, web build, hpc {setup,run,serve}
• Lens workspace (file-based viewer + desktop app): Heatmap, Model, Output,
  Attention, Logit Lens, Activations (TWERA), Attribution (per-head DLA),
  Attribution Graph, Manifold
• Mechanistic interpretability: per-layer + per-head direct logit
  attribution, attribution graphs, activation-manifold geometry (PCA /
  intrinsic dim / periodicity), CSV/JSON/PNG export, activation diff
• 10 example experiments: induction, IOI, recall-probe, successor heads,
  greater-than, multi-hop, attention demo, reasoning, line-wrap helix, cyclic-days ring
• HPC round-trip (hpc run) and a MkDocs documentation site

Early alpha — APIs and the trace schema may still change.