singhpratech/emc2-lesson-02-attention

e=mc² Lesson 02: How Attention Works

Read the attention maps of Tiny Storyteller (Lesson 01, 5.85M parameters, 6 blocks x 4 heads). No training: the notebook loads the public weights, plots heatmaps, and measures all 24 heads over the 200 held-out quiz stories.

Lesson: https://theaivibe.org/emc2/how-attention-works-tiny-storyteller-attention-maps

Run all cells: the widget and figures appear when the cells execute.

Headline numbers (full table in RESULTS.md): the head that puts the most attention on the main character's name is block 5 head 3 (L5H3), with 0.221 mean attention mass on name tokens (about 4.3x its fair share of 0.052); 0 of 24 heads are mostly previous-token heads (mean mass on position i-1 above 0.5; highest 0.146).

Files: lesson_02.ipynb (unexecuted), RESULTS.md, results.json, figures/*.png. Runs in under 20 seconds on an Apple Silicon Mac. Author: Prateek Singh.