SSA Gist Tokens — Unified Visual Deck

This single HTML fuses the three SSA visuals while preserving their original layouts and content as much as possible. It is organized as a presentation-style page: overview, training mechanism, then the full raw-token/gist-token mask example.

1. ConceptGist token = learned chunk summary + deterministic pointer to raw KV range.
2. TrainingGist-aware mask forces future prediction through gist bottlenecks.
3. MaskRows are query tokens; columns are key tokens, including raw-token query rows.
Source visual 1

Chapter 1 — Code-verified overview

Original SSA gist-token visual summary.

Back to top
Source visual 2

Chapter 2 — Training mechanism

Extended explanation of how gist tokens are trained.

Back to top
Source visual 3

Chapter 3 — Full mask example

Three-gist toy example with raw-token and gist-token query rows.

Back to top

Unified from: ssa_gist_token_visual.html, ssa_gist_token_visual_training.html, and ssa_3_gist_attention_visual_full_mask.html.