Results · NN Dataset

The NN Dataset

An LLM-generated dataset of 608 PyTorch neural network implementations, designed to support research on neural network verification, refactoring and migration.

Explore the dataset
608
Neural networks
7
Architecture types
4
Task categories
6,842
Layers and tensor operations
How it was built

Diversity by design

The dataset was generated by systematically varying four key design dimensions, producing neural network implementations with diverse and controlled coverage of real-world configurations.

Architecture type

  • MLP
  • CNN-1D
  • CNN-2D
  • CNN-3D
  • RNN-Simple
  • RNN-LSTM
  • RNN-GRU

Task category

  • Binary classification
  • Multiclass classification
  • Regression
  • Representation learning

Input data type

  • Tabular
  • Time series
  • Text
  • Image

Model complexity

  • Simple
  • Wide
  • Deep
  • Wide-Deep
How it was generated

The generation workflow

The dataset was generated using GPT-5, guided by carefully crafted prompts that ensure diversity across architectures, tasks, input types and scales, and complexity levels.

01

Requirements definition

Each network is characterised by four categories of requirements: architecture type, learning task, input type and scale, and complexity, ensuring diversity and coverage across real-world configurations.

02

Prompt construction

Each combination of requirements is translated into a structured prompt that instructs the LLM to produce a complete, self-contained PyTorch implementation.

03

LLM-based generation

GPT-5 generates the neural network code based on the structured prompt, producing a diverse and controlled set of 608 implementations.

Verified and compliant

Three levels of validation

Each neural network in the dataset was verified through a dedicated validation process. This covers compliance with the specified design dimensions, code correctness and practical utility of the dataset as a whole.

Design compliance

Each network was verified against its four specified design dimensions using static analysis and symbolic tracing, confirming that architecture type, task category, input type and complexity requirements are all satisfied.

Functional validation

The generated networks were trained and evaluated on standard benchmark datasets to confirm their functional correctness across a diverse set of architecture types.

Utility demonstration

The dataset was used to evaluate a neural network migration tool, confirming its value as a benchmark for assessing and improving NN tools.