Heptapod-120M

Heptapod-120M is a 125,829,354-parameter nanochat base model pretrained in a globally base-7 numeral world. Every maximal ASCII decimal digit run in the paired base-10 corpus was rendered in base 7 while document order, arithmetic placement, tokenizer, architecture, seed, optimizer, schedule, and training token budget were held fixed.

The name refers to the seven-fold numeral world and the model's native T7 Fourier phase geometry examined in the accompanying interpretability experiments.

Training configuration

  • Architecture: nanochat depth 8, width 512, 4 attention heads
  • Parameters: 125,829,354 total
  • Context length: 2,048
  • Vocabulary: 32,768
  • Training budget: approximately 2.013B tokens
  • Parameter:data ratio: 48
  • Seed: 42
  • Hardware: 1x NVIDIA H100 SXM 80GB
  • Training precision: FP8 matmuls with master checkpoint weights
  • Final step: 7,680
  • nanochat commit: 92d63d4e8bb4df75c3b71618f31ddde2378b2bcd

Data and tokenizer

The tokenizer is exactly the same 32K digit-level tokenizer used by the paired base-10 baseline. Tokens 0 through 6 represent base-7 digits. Tokens 7 through 9 remain in the unchanged vocabulary but are not valid base-7 digits.

Checkpoint format

These are raw nanochat PyTorch checkpoints, not Transformers AutoModel checkpoints. Reconstruct the model using the pinned nanochat commit above and place the tokenizer/checkpoint directories under NANOCHAT_BASE_DIR.

The repository includes intermediate and final model checkpoints, optimizer states, metadata, training logs, tokenizer files, and the exact RunPod scripts. Evaluation and Fourier/steering outputs can be added by rerunning the uploader after those artifacts are created under evals/heptapod-d8-r48-seed42/.

Paired experiment

This repository is not affiliated with Karpathy, NVIDIA, or Hugging Face.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support