anon-hf-graph-corpus/hf-graph-corpus

Hugging Face Computational Graph Corpus

The Hugging Face Computational Graph Corpus contains computational graph representations extracted from traceable PyTorch models on Hugging Face. Each model is traced through its forward computation, canonicalized, hashed, and represented with graph-level structural features for architecture analysis, duplicate detection, retrieval, and visualization.

The corpus is intended for studying machine learning models as executable architectures rather than only through names, tags, or model-card metadata. It supports research on model ecosystem structure, architectural reuse, graph-based model similarity, and large-scale model cartography.

Dataset Summary

This dataset was built by tracing PyTorch-compatible Hugging Face models using a graph-capture pipeline, then canonicalizing the resulting computational graphs to identify repeated architectures. In the associated analysis, 67,397 models were successfully traced, yielding 1,810 unique graph architectures after deduplication. The high duplicate fraction reflects widespread reuse of common templates, fine-tunes, retrained checkpoints, and architectural variants.

Intended Uses

• Computational graph analysis of neural network architectures
• Architecture-aware model search and retrieval
• Duplicate and near-duplicate architecture detection
• Corpus-level visualization of model families
• Studying architectural reuse across the Hugging Face ecosystem

Limitations

The corpus represents a traceable PyTorch subset of Hugging Face models, not the full Hugging Face model ecosystem. Models may be excluded if they cannot be downloaded, loaded, supplied with suitable example inputs, or traced successfully. Graphs are also dependent on tracing choices, input synthesis, and canonicalization rules, and they do not capture learned weights, training data, safety behavior, or downstream performance.