Datasetpublic

Nozomi1856/NAG-MEQD-Experiment-Data

NAG-ME-QD Experiment Data Trained model weights, elite archive, and full experiment results from the Neural-Architecture-Generation-using-MAP-Elites-Quality-Diversity research project. Code repository: github.com/[Nozomi1856]/Neural-Architecture-Generation-using-MAP-Elites-Quality-DiversityAuthor: Pratheeksha Aravind Experiment configuration Setting Value Dataset CIFAR-10 MAP-Elites iterations 20 Initial population 10 Evaluation mode Zero-shot… See the full description on the dataset page: https://huggingface.co/datasets/Nozomi1856/NAG-MEQD-Experiment-Data.

sourceHugging Facemitupdated 7mo agoView on Hugging Face

0likes5downloads

Dataset Card

NAG-ME-QD Experiment Data

Trained model weights, elite archive, and full experiment results from the Neural-Architecture-Generation-using-MAP-Elites-Quality-Diversity research project.

Code repository: [github.com/[Nozomi1856]/Neural-Architecture-Generation-using-MAP-Elites-Quality-Diversity](https://github.com/Nozomi1856/Neural-Architecture-Generation-using-MAP-Elites-Quality-Diversity) Author: Pratheeksha Aravind

Experiment configuration

Setting	Value
Dataset	CIFAR-10
MAP-Elites iterations	20
Initial population	10
Evaluation mode	Zero-shot
Final training epochs	5
Behavior space dimensions	32
Seed	42

Total runtime: ~24.8 minutes

Results summary

Model	Cell type	Val accuracy	Parameters
top1 (`fe0b192e6185`)	HYBRID	54.4%	40,842
top2 (`705bc1677934`)	HYBRID	56.1%	56,330
top_transformer (`7266c514462f`)	HYBRID	58.0%	57,482
top_recurrent (`268dec31f257`)	DAG	44.3%	77,194
top_conv (`fe0b192e6185`)	HYBRID	53.8%	40,842

Archive: 24 elites, QD score 11.99, max fitness 0.612, mean fitness 0.500

File structure

nag_output/
├── final_results.json              # Full experiment summary and config
├── final_qd_trajectory.json        # QD score, coverage, fitness per iteration
├── final_archive_changes.json      # Per-iteration archive update history
├── final_mutation_stats.json       # Mutation operator success rates
├── behavior_space_reference.json   # 32D behavior dimension definitions
├── qd_trajectory_plot.png          # Fitness/coverage plot
│
├── top1/
│   ├── architecture.json           # Architecture specification
│   ├── model.pth                   # Trained PyTorch weights
│   └── training_history.json       # Per-epoch loss and accuracy
├── top2/                           # (same structure)
├── top_conv/                       # (same structure)
├── top_transformer/                # (same structure)
├── top_recurrent/                  # (same structure)
│
└── elite_archive/
    ├── elite_index.json            # Index of all 24 elites with fitness and tags
    └── elite_[id].json             # Per-elite architecture spec and behavior vector

Loading a model

python

import torch
import json

# Load architecture spec
with open("nag_output/top_transformer/architecture.json") as f:
    arch_spec = json.load(f)

# Load weights
weights = torch.load("nag_output/top_transformer/model.pth", map_location="cpu")

# To rebuild the model, use NASNetwork from the code repo:
# https://github.com/Nozomi1856/Multi-Paradigm-MAP-Elites-Quality-Diversity-Neural-Architecture-Generation-with-DARTS-cells

Notes

The zero-shot fitness scores used during search are proxy estimates, not true validation accuracy. Final accuracy figures above are from short post-search training runs (5 epochs). Results likely reflect basic CNN structure and standard training hyperparameters more than the NAS components — this is a known limitation of the current implementation.

Citation

If you use this data, please credit:

Pratheeksha Aravind. NAG-ME-QD: Neural Architecture Generation with MAP-Elites 
Quality-Diversity Optimization. Work in progress, 2026.
GitHub: https://github.com/Nozomi1856/Neural-Architecture-Generation-using-MAP-Elites-Quality-Diversity/settings