# Numbat: Building and Verifying a Self-Contained Machine-Learning Stack

arXiv: https://arxiv.org/abs/2609.10632 · HTML: https://mlsanity.com/paper/2609.10632/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 26 of its 40 references that are in the map.
- First posted on arXiv 9 Sep 2026.

## Builds on (most cited first, 26 of 26)

- [PyTorch: An Imperative Style, High-Performance Deep Learning Library](https://mlsanity.com/paper/1912.01703/) (2019)
- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift](https://mlsanity.com/paper/1502.03167/) (2015)
- [mixup: Beyond Empirical Risk Minimization](https://mlsanity.com/paper/1710.09412/) (2017)
- [Feature Pyramid Networks for Object Detection](https://mlsanity.com/paper/1612.03144/) (2016)
- [You Only Look Once: Unified, Real-Time Object Detection](https://mlsanity.com/paper/1506.02640/) (2015)
- [Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification](https://mlsanity.com/paper/1502.01852/) (2015)
- [YOLOv3: An Incremental Improvement](https://mlsanity.com/paper/1804.02767/) (2018)
- [YOLOv4: Optimal Speed and Accuracy of Object Detection](https://mlsanity.com/paper/2004.10934/) (2020)
- [Random Erasing Data Augmentation](https://mlsanity.com/paper/1708.04896/) (2017)
- [Mixed Precision Training](https://mlsanity.com/paper/1710.03740/) (2017)
- [YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors](https://mlsanity.com/paper/2207.02696/) (2022)
- [Sigmoid-Weighted Linear Units for Neural Network Function Approximation in Reinforcement Learning](https://mlsanity.com/paper/1702.03118/) (2017)
- [Path Aggregation Network for Instance Segmentation](https://mlsanity.com/paper/1803.01534/) (2018)
- [Deep Reinforcement Learning that Matters](https://mlsanity.com/paper/1709.06560/) (2017)
- [Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation](https://mlsanity.com/paper/2012.07177/) (2020)
- [Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression](https://mlsanity.com/paper/1911.08287/) (2019)
- [Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection](https://mlsanity.com/paper/2006.04388/) (2020)
- [CSPNet: A New Backbone that can Enhance Learning Capability of CNN](https://mlsanity.com/paper/1911.11929/) (2019)
- [Rethinking ImageNet Pre-training](https://mlsanity.com/paper/1811.08883/) (2018)
- [Improving Reproducibility in Machine Learning Research (A Report from the NeurIPS 2019 Reproducibility Program)](https://mlsanity.com/paper/2003.12206/) (2020)
- [YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications](https://mlsanity.com/paper/2209.02976/) (2022)
- [TOOD: Task-aligned One-stage Object Detection](https://mlsanity.com/paper/2108.07755/) (2021)
- [Fast Algorithms for Convolutional Neural Networks](https://mlsanity.com/paper/1509.09308/) (2015)
- [DSOD: Learning Deeply Supervised Object Detectors from Scratch](https://mlsanity.com/paper/1708.01241/) (2017)
- [Cross-Stack Validation of Language-Model Training: A Clinical Fine-Tuning Case Study](https://mlsanity.com/paper/2608.24267/) (2026)

## Built on by (most cited first, 1 of 1)

- [woma: a real-time foundation model and its fine-tuned models for endoscopy](https://mlsanity.com/paper/2609.15130/) (2026)

## Abstract (arXiv)

Machine-learning systems are built almost exclusively on a few large Python-orchestrated frameworks, and they inherit those stacks' engineering costs: environments of hundreds of version-coupled packages, separate export toolchains for deployment, and the split between the language research is written in and the language products ship in. We report on the construction and verification of numbat, a machine-learning stack written in one general-purpose language (Zig) with no third-party runtime dependencies. The stack spans tensor computation, automatic differentiation, neural-network modules, mixed precision, multi-GPU training, data loading and monitoring; an SDK exposes it behind a stable, additively versioned C ABI of over 1,400 entry points, with bindings for six languages; and its clinical domain planes encode regulatory requirements as executable acceptance gates rather than documentation. Verifying such a stack is the harder half of building it: a defective training run rarely fails, it converges quietly to a slightly worse model. We treat a widely used reference implementation as an executable specification and verify against it at five levels, from operator gradient checks to an automated trajectory gate against a same-machine reference run - the arrangement our companion study formalizes as a trajectory-level differential oracle. The protocol surfaced ten silent recipe divergences, which we catalog with mechanisms and symptoms. As the acceptance test, we train a 25.9M-parameter detector of the YOLOv8m class from random initialization on COCO 2017 for the full 500-epoch schedule: the exported weights score 0.4956 mAP50-95 under the official protocol, scored by the reference stack's own validator (published endpoint 0.502), with single-GPU step time at parity on identical hardware. Weights, per-epoch metrics and the full run manifest are released.