# tysam-code/hlb-CIFAR10

Train to 94% on CIFAR-10 in <6.3 seconds on a single A100. Or ~95.79% in ~110 seconds (or less!)

Repository: https://github.com/tysam-code/hlb-CIFAR10
Canonical: https://ross.abutalabs.com/products/hlb-cifar10
Homepage: https://twitter.com/hi_tysam
Language: Python
License: Apache-2.0
License Family: permissive
Topics: cifar10, deep-learning, experimentation, machine-learning, world-record, language-models-are-next, responsive-to-issue-tickets, neural-network, speedyconvnet7-1-0
Last push: 2024-12-18T03:12:35+00:00

## Health v2 (maintenance only)
Score: 22/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 8, longevity 97
- inputs: {"age_days": 1361, "days_push": 623, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1310, forks 81 (observed 2026-08-28T04:04:19.836638+00:00)

## What it is
A single-file PyTorch implementation that trains a neural network to 94% accuracy on CIFAR-10 in under 6.3 seconds on a single A100 GPU, formerly holding the single-GPU training speed world record. It is designed to be minimal, hackable, and beginner-friendly for rapid deep-learning experimentation.

## Use cases
- train a CIFAR-10 classifier in seconds on one GPU
- rapidly experiment with neural network training ideas
- reproduce ultra-fast CIFAR-10 training results
- benchmark single-GPU deep learning training speed
- learn how fast CIFAR-10 training pipelines work
- hack on a minimal deep learning training codebase

## When to choose
- you want record-fast CIFAR-10 training on a single GPU
- you need a small, flat, hackable codebase for training experiments
- you want a beginner-friendly PyTorch training reference
- you're working in Colab or a CUDA environment

## When to avoid
- you need production-grade, maintainable training code
- you don't have access to a CUDA GPU
- you need multi-GPU or distributed training
- you want a configurable framework rather than a single-file script

## Facets
- artifact type: library
- maturity: maintenance
- function: deep-learning, machine-learning, benchmarking
- domain: deep-learning, machine-learning, computer-vision, performance
- platform: python, cloud
- tags: cifar10, fast-training, pytorch, single-file, research-code, record-setting, colab-friendly, gpu, linux

## Member repositories
- tysam-code/hlb-CIFAR10 (main) score 22

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:19.836638+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:49:57.038882+00:00, confidence not recorded.
  - readme: https://github.com/tysam-code/hlb-CIFAR10 (fetched 2026-08-28T04:04:19.836638+00:00, sha 9dcef5ee086f)
- Data as of 2026-08-30T08:39:29.467469+00:00.