README.md

metadata

license: mit
tags:
  - mechanistic-interpretability
  - grokking
  - transformers

Manifold Features — data

State dicts and supporting checkpoints for the Manifold Features interactive longform series by Nikolay Yudin.

The source code and notebooks live at github.com/nick-yudin/manifold_features.

Contents

four-algorithms/ — Part 2 checkpoints

A 2-block transformer with residual dimension 384 trained on four modular-arithmetic tasks simultaneously (division, addition, max, parity at P = 149) with WD = 0.3 across 5 random seeds.

• seed_1000_final.pt ... seed_1004_final.pt (~19 MB each) — final checkpoints after grokking; include the model state dict, the metrics log, the per-task grok steps, and the snapshot step list.
• handcraft_max.pt (~18 MB) — a 2-block transformer that solves c = max(a, b) mod 97 with weights placed entirely by formula (no training). Hits 100% accuracy on all 9409 input pairs.

Loading

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download(
    repo_id="NikolayYudin/manifold-features-data",
    filename="four-algorithms/seed_1000_final.pt",
    repo_type="dataset",
)
ckpt = torch.load(path, map_location="cpu", weights_only=False)
print(ckpt["final_accs"])  # {'div': 0.9995, 'add': 0.9854, 'max': 0.9915, 'parity': 0.9995}

Architecture matches GrokModelShared from the notebooks. Open four-algorithms/notebooks/walk_and_dictation.ipynb for the full loading + analysis flow.