alife2026-model

ALife 2026 model content

This directory is excluded from the single GitHub repository rooted at alife2026/. It is the local content tree managed and published through the ALife 2026 Hugging Face model repository, and stores trained PyTorch controllers separately from source code and experimental datasets.

Layout

Paths preserve their former experiment-relative structure:

model/
├── best_brain.pt
└── <experiment>/
    └── models/
        ├── <condition>_run<N>.pt
        └── checkpoints/

The primary frozen evaluation uses:

v4.4_cross_20260224_2238_G50000_R3_C5000/models/

It contains six conditions, three matched training runs per condition, and two controller architectures (MLP4 and Wide128). In the complete local workspace, the authoritative SHA-256 list is tracked by the root GitHub repository at:

doc/evaluation/linux_eval_handoff/MODEL_SHA256SUMS.txt

Loading

import torch
from pathlib import Path

# Run from the alife2026 workspace root.
path = Path("model") / (
    "v4.4_cross_20260224_2238_G50000_R3_C5000/"
    "models/cross_evoevo_mlp4_run0.pt"
)
checkpoint = torch.load(path, map_location="cpu", weights_only=True)

The source code under src/ discovers this directory through ALIFE_MODEL_ROOT, which defaults to the root-level model/ directory.

The frozen-evaluation source baseline before this workspace reorganization was commit 520c409b8e5f34403c4a13a8308e626cbeb44a4b. Record the final root-repository commit alongside any newly published evaluation batch.