NeWGen-Bench

A benchmark for neural network parameter generation. The release contains trained checkpoints across multiple architectures, datasets, and task domains, each paired with structured metadata describing the architecture, training data, hyperparameters, and final performance.

This dataset accompanies an anonymous submission to NeurIPS 2026 Evaluations and Datasets Track.

Repository Structure

<root>/
  datasets/   Local datasets and dataset caches
  registry/   JSON registries for benchmark runs, metrics, configs, and checkpoints
  train/      Training scripts and configuration files
  models/     Model definition files
  zoo/        Trained model checkpoints
  collect/    Collected checkpoint trajectories

Task Domains

The benchmark spans the following task domains:

• Image classification
• Image segmentation
• Text classification
• LLM adaptation
• VLM adaptation

Artifacts

• Trained checkpoints: zoo/**/*.pth
• Collected checkpoint trajectories: collect/**/*.pth
• Training configs: train/**/config/*.json
• Model definitions: models/*.py

Status

This is an initial placeholder release. The metadata table and a representative sample of checkpoints will be added before the full paper deadline. The complete collection of trained checkpoints will be made available before the camera-ready deadline.

License

Released under CC BY 4.0. Individual checkpoints are derived from public training datasets, each governed by its own license; downstream users should comply with the licenses of the underlying training data.