Graph classification just got more accessible. Hugging Face’s Transformers library now supports Microsoft’s Graphormer, currently the only graph transformer model available in the ecosystem, and the team has published a practical walkthrough for fine-tuning it on molecular data. The tutorial centers on ogbg-molhiv, a baseline dataset from Stanford’s Open Graph Benchmark, which loads with a single line: load_dataset("OGB/ogbg-molhiv"). The dataset arrives pre-split into train, validation, and test sets, each containing five columns — edge_index, edge_attr, y, num_nodes, and node_feat.
What makes this tutorial genuinely useful is its clarity around data formatting. Graph datasets on the Hub are stored as lists of dictionaries, and the walkthrough spells out exactly what each field means. The edge_index field holds two parallel lists of node indices representing connections, while num_nodes clarifies the total node count — critical because isolated nodes don’t appear in edge lists. The y field handles multi-class, regression, or multi-task binary labels, and optional node_feat and edge_attr fields capture node types and edge attributes, like atom types or molecular bonds.
Preprocessing is where Graphormer does its heavy lifting. The default pipeline generates in/out degree information and shortest-path matrices between nodes — features that help the transformer understand graph topology. The tutorial shows two approaches: preprocessing the entire dataset upfront with preprocess_item, or setting on_the_fly_processing=True in the collator for larger graphs where storing preprocessed data would be prohibitively expensive. That flexibility matters for real-world use cases beyond toy benchmarks.
The model itself is loaded from a pretrained checkpoint — clefourrier/pcqm4mv2_graphormer_base — and fine-tuned for binary classification with num_classes=2. The ignore_mismatched_sizes=True flag handles the architecture adjustment. What’s notable here is the underlying bet: pretraining on PCQM4Mv2, a massive quantum chemistry dataset, should transfer well to molecular property prediction tasks like HIV inhibition. Whether that transfer actually beats simpler graph neural networks on small datasets like ogbg-molhiv remains an open question — the tutorial doesn’t benchmark against GCNs or GINs, which have historically dominated OGB leaderboards. Still, for practitioners already invested in the Hugging Face ecosystem, having graph transformers work with the same from_pretrained and trainer APIs as text and vision models lowers the barrier considerably.
