Advanced NLP assignment 3: Part B
Forty from-scratch PyTorch GossipCop classifiers: five seeds for each of eight unique architecture/feature configurations. GraphSAGE BERT is shared between the architecture and feature comparisons. BiGCN follows Chen et al., arXiv:2101.05519v2, using node-and-feature low-pass filtering. Superseded Bian et al. checkpoints are excluded.
Each checkpoint was selected by highest validation macro F1, with earliest-epoch tie handling. Results use the official UPFD 20/10/70 splits and macro precision, recall, and F1. results.json contains means; standard-deviations.json contains five-seed sample standard deviations. These are measured test results, not validation-selection scores.
Load weights.safetensors and metadata.json using the assignment repository's src.part2.models.GraphClassifier and the model_settings in each metadata file. The manifest maps every configuration and seed to its checkpoint folder and original W&B run ID. AutoModel integration is not provided. Code is supplied separately through the assigned course repository. Raw data, optimizer states, credentials, and machine-specific paths are excluded. This card documents model artifacts and is not the student's assignment report.