affordable-frontier-architectures: Hugging Face mirror with checkpoints
This repository mirrors the parts of the affordable-frontier-architectures research repository that include model checkpoint files.
- Source of truth (code, reports, results, checkpoints): https://github.com/Marxist-Leninist/affordable-frontier-architectures
- Reports, code and results on Hugging Face: https://huggingface.co/MarxistLeninist/affordable-frontier-architectures
Both accounts are maintained by the same author. Checkpoint files are hosted under this account because the public storage quota of the MarxistLeninist account is exhausted, which blocks new large-file uploads there.
Contents
experiments/nat_transformer_mamba_20261111/ is a complete copy of the public release of the 800-update single-pass NAT Transformer versus bidirectional Mamba-1 pilot: code, protocols, metrics, audits and all 16 small FP32 checkpoints. Every file in that directory is byte-identical to GitHub commit 1acf25d54bdcbf532ade1ff827b60b9a5583e138.
experiments/nat_transformer_mamba_20261111/FILE_MANIFEST.json lists the SHA256 of every file in the directory, including the 16 checkpoints in nat_compare_20261111/checkpoints/.
Start with experiments/nat_transformer_mamba_20261111/README.md and nat_compare_20261111/REPORT.md for what was measured and the limits of the result.
About the checkpoints
They are tiny research artifacts, not useful conversational models: 94,416 parameters for the Transformer and 94,320 for the Mamba reference. Each file is a plain PyTorch state dict. Load one with torch.load(path, weights_only=True, map_location="cpu") and the NAT class in nat_compare_20261111/models.py.
This mirror does not assign a license to the code or checkpoints. See nat_compare_20261111/THIRD_PARTY_NOTICES.md for third-party attribution.
Gallery screen and independent verification (12 November 2026)
experiments/gallery_cpu120_20261110/ contains the complete historical 105-record ASI-Arch screen, its code, 146 result files and all six retained checkpoints. This is an adapted approximately 350k-parameter, 120-CPU-second experiment, not a full-scale architecture winner.
Independent verification and reproduction repairs documents the repaired fresh-checkout workflow and exact input/state hash checks. None of the six checkpoint reloads meets the original strict loss tolerance on a different CPU/PyTorch version; that portability limitation is preserved. Fifteen key files including all six checkpoints were downloaded from the initial gallery upload and verified by SHA-256. No new training was run for that publication.