kyr0/evoweights-evolutionary-weight-space-fine-tuning

EvoWeights pre-trained models

Arena: kyr0.github.io/evoweights - Code: github.com/kyr0/evoweights

These are the models shipped in the EvoWeights browser arena. All of them are NVIDIA Kumo-Tabular Small (27.46 M parameters), an in-context tabular classifier. The classifier picks one of three actions (LEFT, RIGHT, FIRE) from game observations and 96 labelled demonstration rows. 119 signed sensor weights turn live game sensors into pushes on the three scores and add movement (see the paper, section 03).

Layout

Path	Content
base/fp32/	Unchanged FP32 checkpoint, gzip byte-plane packs plus original-manifest.json (≈ 79 MB)
base/4bit/	4-bit quantized base with the 96-row support cache, packs plus manifest.json (≈ 20 MB)
revisions/*.json	One file per model: the 4-bit code changes (ops) and sensor weights (controller) applied to a base
LICENSES/	OpenMDW-1.1 for the Kumo data, Apache-2.0 and BSD-3-Clause notices for the runtime
SHA256SUMS	Checksums of every file in this folder

A model is its base plus its revision file. The id in each revision file is the SHA-256 revision ID the arena derives from {precision, parentId, ops, controller}, so a changed file gets a different ID.

Models

Reference total: tics survived, summed over the arena's 48 fixed reference games (10,500-tic cap each), measured in Node.js. Factor and standard error are relative to Original FP32.

Model	Revision file	Parent	How it was made	Reference total	vs FP32
Base Model FP32	original-fp32.json	–	No training	22,564	1.00×
Astra 6 Weight-Space Programming	astra6-weight-space-programming.json	FP32 → 4-bit	OpenAI Astra 6 set 256 FIRE-row codes and the sensor weights directly	65,150	2.89× ± 0.25
4-bit base · no FIRE edit	base-4bit-no-fire-edit.json	FP32 → 4-bit	Training root: Astra 6 sensor weights without the FIRE edit	64,856	2.87× ± 0.26
4-bit base · neutral sensor weights	base-4bit-neutral-sensor weights.json	FP32 → 4-bit	Training root: every sensor weight 0	22,943	1.02× ± 0.09
EvoWeights 4-bit · R1	evoweights-r1.json	no FIRE edit	EvoWeights protocol 7 in Node.js	98,447	4.36× ± 0.38
Opus 5.5 Weight-Space Programming	opus-weight-space-programming.json	neutral sensor weights	Claude Opus 5.5 set the sensor weights without access to earlier values	88,570	3.93× ± 0.97
EvoWeights Survivor · R3	evoweights-survivor.json	neutral sensor weights	EvoWeights in the browser on CPU, survival metric	295,358	13.09× ± 2.21
EvoWeights Terminator · R4 · Kills	evoweights-terminator.json	neutral sensor weights	EvoWeights in the browser on CPU, kills metric	82,401	3.65× ± 0.36

The reference total measures survival only. The Terminator was selected for kills, so its total understates what it was trained for. The Survivor's total measured in Chrome was 280,967 (+1,145 %), 5.1 % below the Node.js value; the cause is UNKNOWN (experiments/v11/edno-neutral-browser/RESULTS.md). Held-out results, paired t-tests and losses are in the paper.

Use

Open the arena, choose a model card and press Play, or run the Node.js tools of the repository:

Command	What it does
make setup	bun install --frozen-lockfile, the Playwright Chromium and the model packs (make models)
make models	downloads the model packs from Hugging Face into .cache/model/ and checks each against its manifest; makes no request when verified copies exist
make build	compiles src/ and copies public/ into docs/, the folder GitHub Pages deploys
make start / make stop	builds, then serves docs/ on http://localhost:8080/ with the cross-origin isolation headers (logs in var/log/)
make test	unit and integration tests against the shipped WASM, the local pack copy and real workers (bun test)
make coverage	the same with line coverage
make lint	tsc for src/ and for tests and scripts, oxlint, prettier --check
make e2e	builds and drives every page section in Chromium; the report goes to output/
make verify	lint, test, coverage and e2e

Rebuilding the WASM binaries from source is described in source/README.md.

Sources and credit

• Base model: nvidia/Kumo-Tabular, small classifier v1.0.0. The 4-bit parameters and derived key/value tensors are distributed under OpenMDW-1.1.
• EvoWeights method, arena and paper: Aron Homberg.
• Agent roles: OpenAI Astra 6 wrote the scripted demonstrator, the FIRE-row edit and the shipped sensor weights. Claude Fable 5.1 created the 4-bit training roots. Claude Opus 5.5 wrote the EvoWeights implementation and made the independent edit.