NM Project indexCode and recorded results
All projects

Integer & embedded · Research

int-llm visualizations

Four offline browser views of an integer GPT, with exact values and C-oracle checks behind the displayed numbers.

View source Open live demo

At a glance

Evidence and scope

Repository measurements and recorded project results. Test and benchmark claims were inspected, not rerun during this catalog refresh. Physical line counts use the stated file scopes.

Visual surfaces
4 offline pagesWeight atlas, inference trace, residual-space viewer, and 3-D weight explorer; data is embedded.
Checkpoint scope
14,272 weightsAll nine learned tensors in the one-layer, 32-dimensional MicroGPT checkpoint.
Exact trace
122 stepsRecorded logits, cache state, samples, and final RNG state are checked against the C runtime.
Layout comparisons
4 neuron ordersOriginal, smooth, spread, and seeded-random MLP layouts, plus a separate residual-coordinate swap.
Reviewed sources · 2026-10-15 · 505e65235871

43 tracked files · 2 documentation files · 0 workflow files. Inventory includes data and vendored or generated files; workflow presence does not establish a passing run.

Download inventory and line-count file scopes

Executive summary

int-llm-viz makes a complete small integer GPT inspectable in a browser. Four self-contained pages show its weights, recorded inference, shared residual space, and weight contributions. Each page embeds its data and works offline. Model values use BigInt arithmetic and exact decimal readouts; floating-point geometry is labeled separately. The C runtime provides the reference trace, and browser calculations check that trace instead of displaying an unchecked animation. The result connects the integer-computing and model-layout projects through visible, testable examples.

What was built

  • An atlas of all nine weight tensors, including exact rounding previews.
  • A step-by-step replay of 20 generated names with browser recomputation.
  • A view of 62 vectors in the 32-dimensional residual space, with literal axes

    • and explicitly derived principal-component projections.

  • A weight explorer combining 3-D geometry, exact per-weight contributions,

    • byte layout, and four-dimensional projections.

  • Reconstructed MLP layouts and a residual-coordinate swap, each replayed

    • through the C oracle.

Evidence and validation

The trace harness reports 122 steps with zero logit, key/value-cache, or sampling mismatches and an identical final random state. Browser tests check the same arithmetic, model-loader checksums, and exact decimal formatting. The release procedure also requires deterministic data and page regeneration, browser interactions, and no drift from committed artifacts.

The four neuron orders and hidden-coordinate swap are rebuilt and checked against pinned hashes. Equal output is established through model replay, not inferred from how similar the pictures appear.

Boundaries

This is a visualization for one checkpoint and shape, not a general model inspection framework. Principal components are derived geometry; they do not establish semantic meanings for coordinates. Embedding all data into each page is practical at this scale but would not scale to a large language model.

Repository map and first run

Open the live demo or the committed site/index.html directly in a browser. To regenerate and validate locally:

make
make test

make release-check additionally requires a browser and fails if regeneration changes the committed data or pages. SOURCE_PROVENANCE.md identifies copied components, upstream pins, and the independent implementation process.

Continue exploring

Integer & embeddedint-llm

Pure-integer C experiments for deterministic language-model training and inference, validated across desktop, embedded, and emulated targets.

Integer & embeddedint-llm coordinate permutation

A toolkit for controlled model-layout experiments on both standard SafeTensors and the integer Q16.48 model path.

Integer & embeddedint-llm precision ladder

Follow-up experiments that narrow integer model weights while retaining the wide Q16.48 path as the comparison oracle.

Integer & embeddedint-llm SQL

SQL executes model loading, transformer arithmetic, deterministic sampling, and regression checks without custom functions or extensions.