Records · kv:2610.00013v1

Reproduction of tinygrad/tinygrad: hlb_cifar10 and beautiful_mnist examples

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-07

Independent reproduction · Benchmark or leaderboard entry · Machine learning

Reproduction of: original code

Abstract

Ran two training examples from tinygrad at the pinned commit, unmodified and with their default settings, on the NV backend: examples/hlb_cifar10.py (the hlb-CIFAR10 speedrun port, 1000 steps at batch size 512, fixed seed 201 from the script hyperparameters, test accuracy evaluated by the script) and then examples/beautiful_mnist.py (small convolutional network, Adam, 70 steps at batch size 512, test accuracy at the end). Training loss every 50 steps, CIFAR-10 test accuracy and the final MNIST loss and test accuracy are parsed from the script output. CIFAR-10 and MNIST fetched by tinygrad before the sealed run; CUDA NVRTC 12.8 from conda-forge as the kernel compiler. Single NVIDIA RTX 5060 Ti 16 GB, Python 3.12.

Sealed report

Cite this record

@misc{kveritas261000013v1,
  title={Reproduction of tinygrad/tinygrad: hlb_cifar10 and beautiful_mnist examples},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00013v1},
  url={https://kveritas.org/records/2610.00013v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.