Multi-head attention comparison without head averaging
three.js and WebGL · 3D · MIT
- What it does
Compare two to eight supplied attention heads from their separate projections through concatenation and the output projection while retaining each head identity. A written recipe (abstract-systems-data, linear-algebra; part sourced, part authored): the parts to model, the one state that drives them, what must stay true, how it should move and its known limits. Text for an agent to build from, not code.
- Get it
python3 scripts/resolve.py recipe.multi-head-attention-comparison
- Needs (the whole pack)
Nothing to read the skill. Python 3 for the catalog helpers. Node with pnpm or npm for the starter and the examples. Optional: Playwright for its capture script, and Blender 4.2 or newer for its baked hero tier. The six anatomy studies need 169 MB of meshes fetched separately.
- License
MIT
- From