Chemical-Family Shift in a Pretrained ALIGNN Formation-Energy Model - Archival Artifacts

Archival artifacts accompanying the study 'Is Fine-Tuning Worth It? Reusing a Pretrained AI Model to Predict the Stability of New Materials' (under review). Contents: (1) checkpoint_300.pt - a mirror of the public JARVIS-DFT ALIGNN v1 formation-energy pretrained checkpoint (Choudhary 2021; original: figshare record 17005681), provided for reproducibility resilience - the original model terms apply. (2) dataset_structures.tar.gz - dataset structure bundle. The 720 fine-tuned and from-scratch checkpoints are intentionally NOT deposited here: they are fully regenerable from the committed configs, fixed seeds, and training scripts in the repository. Code, data manifests, and full results: https://github.com/TheArchitect999/ALIGNN-domain-shift