Skip to content

Improve Section 04 nanoTabICL visualizations #6

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jxudata merged 2 commits into from
Oct 5, 2026
Merged

Improve Section 04 nanoTabICL visualizations #6

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c97a61e
Align architecture schematic arrows with TabICLv2 core data flow
jxudata Oct 5, 2026
438aff2
Clarify final-block output selection across implementations
jxudata Oct 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions materials/website/index.html
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1068,8 +1068,8 @@ <h3>How does the frozen graph use context to answer a query?</h3>
</div>
<div class="prose reveal">
<p>
Inference repeats the same preprocessing and 3-column / 3-row / 12-ICL-block
forward graph, but stops after prediction. Context labels are embedded; the query
In this classifier-core schematic, inference repeats the nanoTabICL preprocessing
and 3-column / 3-row / 12-ICL-block forward graph, but stops after prediction. Context labels are embedded; the query
label is unavailable; pretrained weights remain frozen. There is no loss,
backpropagation, or task-specific optimizer step in this diagram.
<a class="cite" href="https://arxiv.org/abs/2602.11139" target="_blank" rel="noopener">[Qu et al., 2026]</a>
Expand All @@ -1089,7 +1089,7 @@ <h3>How does the frozen graph use context to answer a query?</h3>
<div>
<span class="mono">same graph · inference mode · weights frozen</span>
<h3 id="inference-live-title">Trace the same toy table without an update step</h3>
<p>This is a shape-faithful schematic. Select a stage to inspect what is reused from training and what changes at inference.</p>
<p>This schematic shows the classifier core. The full browser classifier additionally preprocesses and ensembles eight views. Select a stage to inspect the core data flow.</p>
</div>
</header>
<nav class="inference-live__steps architecture-flow" aria-label="Inference architecture stages">
Expand Down
62 changes: 38 additions & 24 deletions materials/website/js/site.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1389,16 +1389,18 @@ function createTensorScene(canvas, render) {
}
}

function datasetAttention(x, y, labels = null, showOutput = true) {
const queryY = y + 9 * 31;
function datasetAttention(x, y, labels = null) {
// Show the final block explicitly. C′ and Q′ have already passed through
// blocks 1–11, where both context and query read context K/V only.
const queryY = y + 9 * 27;
for (let row = 0; row < 10; row++) {
const cy = y + row * 31;
const cy = y + row * 27;
ctx.fillStyle = row === 9 ? "rgba(255,132,108,.28)" : "rgba(154,124,255,.24)";
ctx.strokeStyle = row === 9 ? TECH.coral : TECH.violet;
ctx.fillRect(x, cy, 72, 20);
ctx.strokeRect(x, cy, 72, 20);
const label = labels?.[row] ?? row % 2;
mono(row === 9 ? "Q · y=?" : `C${row + 1} · y=${label}`, x + 36, cy + 10, 10,
mono(row === 9 ? "Q′ · no y" : `C${row + 1}′ · y=${label}`, x + 36, cy + 10, 10,
row === 9 ? TECH.coral : TECH.text, "center", 650);
if (row < 9) {
arrow(x + 76, cy + 10, x + 150, queryY + 10, "rgba(53,217,255,.52)", 1);
Expand All @@ -1410,19 +1412,12 @@ function createTensorScene(canvas, render) {
ctx.arc(x + 166, queryY + 10, 15, 0, Math.PI * 2);
ctx.fill();
ctx.stroke();
text("Q", x + 166, queryY + 10, 13, TECH.coral, "center", 750);
if (showOutput) {
arrow(x + 183, queryY + 10, x + 222, queryY + 10, TECH.coral, 1.5);
cube(x + 228, queryY - 18, 66, 55, 10, "head", "1×10", TECH.coral);
arrow(x + 306, queryY + 10, x + 336, queryY + 10, TECH.gold, 1.5);
ctx.fillStyle = "rgba(255,209,102,.22)";
ctx.strokeStyle = TECH.gold;
ctx.fillRect(x + 342, queryY - 17, 55, 23);
ctx.fillRect(x + 342, queryY + 13, 34, 23);
ctx.strokeRect(x + 342, queryY - 17, 55, 23);
ctx.strokeRect(x + 342, queryY + 13, 34, 23);
mono("2 classes", x + 370, queryY + 52, 11, TECH.gold, "center", 700);
}
text("Q″", x + 166, queryY + 10, 13, TECH.coral, "center", 750);
// The query's own representation participates in attention and residuals.
// Its label is absent, but its features must not disappear from the graph.
arrow(x + 76, queryY + 10, x + 149, queryY + 10, TECH.coral, 1.5);
mono("block 12 shown · query output retained", x + 115, queryY + 37, 9, TECH.muted, "center", 700);
return { outputX: x + 183, outputY: queryY + 10 };
}

function background() {
Expand Down Expand Up @@ -1541,17 +1536,28 @@ function initTrainingLive() {
arrow(635, 267, 662, 267, TECH.violet, 2);
panel(669, 82, 271, 372, TECH.coral);
mono("3 · DATASET ICL ATTENTION", 687, 104, 12, TECH.coral, "left", 700);
mono("11 blocks: C,Q read C · final: Q reads C", 687, 124, 9, TECH.muted);
datasetAttention(692, 146, y, false);
arrow(875, 435, 898, 435, TECH.coral, 1.4);
mono("LN→MLP→1×10", 934, 450, 9, TECH.gold, "right", 700);
mono("before block 1: C += y_embed_icl · Q has no y", 687, 124, 9, TECH.coral);
mono("blocks 1–11: C,Q read C K/V → C′,Q′", 687, 144, 9, TECH.muted);
const iclOutput = datasetAttention(692, 163, y);
arrow(iclOutput.outputX, iclOutput.outputY, 925, iclOutput.outputY, TECH.coral, 1.4);
mono("LN + MLP", 906, 395, 8, TECH.gold, "center", 700);
mono("1×10", 906, 432, 9, TECH.gold, "center", 700);
} else if (active === 2) {
title("Step 03 · query loss", "The toy query contributes 10-way cross-entropy", "illustrative values · training only");
panel(72, 115, 260, 240, TECH.coral);
mono("HIDDEN TRUTH", 202, 144, 12, TECH.coral, "center", 700);
text(`query y = ${y[9]}`, 202, 218, 30, TECH.text, "center", 800);
mono("query label never enters forward pass", 202, 319, 11, TECH.muted, "center");
arrow(350, 235, 405, 235, TECH.coral, 2);
// The hidden label is a loss target, never an input to the output head.
ctx.strokeStyle = TECH.coral;
ctx.lineWidth = 1.5;
ctx.beginPath();
ctx.moveTo(202, 355);
ctx.lineTo(202, 382);
ctx.lineTo(841, 382);
ctx.stroke();
arrow(841, 382, 841, 357, TECH.coral, 1.5);
mono("hidden target → loss only", 490, 370, 11, TECH.coral, "center", 700);
panel(426, 115, 250, 240, TECH.cyan);
mono("ILLUSTRATIVE OUTPUT", 551, 144, 12, TECH.cyan, "center", 700);
cube(492, 181, 118, 82, 15, "LN + MLP", "1×10 logits", TECH.cyan);
Expand Down Expand Up @@ -1673,8 +1679,16 @@ function initInferenceLive() {
} else {
title("Stage 04 · dataset ICL + output", "Twelve ICL blocks turn context into query logits", "no loss · no backward · no update");
panel(49, 90, 474, 356, TECH.coral);
mono("C += y_embed_icl · Q has no y · K/V = context only", 286, 116, 10, TECH.coral, "center", 700);
datasetAttention(82, 139, y, false);
mono("before block 1: C += y_embed_icl · Q has no y", 286, 116, 10, TECH.coral, "center", 700);
mono("blocks 1–11: C,Q read C K/V → C′,Q′", 286, 140, 10, TECH.muted, "center", 700);
const iclOutput = datasetAttention(82, 159, y);
v.ctx.strokeStyle = TECH.coral;
v.ctx.lineWidth = 1.5;
v.ctx.beginPath();
v.ctx.moveTo(iclOutput.outputX, iclOutput.outputY);
v.ctx.lineTo(550, iclOutput.outputY);
v.ctx.lineTo(550, 268);
v.ctx.stroke();
arrow(550, 268, 603, 268, TECH.coral, 2);
panel(628, 134, 268, 270, TECH.gold);
mono("FROZEN OUTPUT HEAD", 762, 163, 12, TECH.gold, "center", 700);
Expand Down
Loading