These are example end-of-run model weights, one per architecture: LSTM, DLinear, PatchTST, iTransformer, and DCRNN. They are reference checkpoints rather than a per-seed canonical set, because the training pipeline overwrites models/ on each run, so each file is just the last training of that architecture from whichever sweep finished most recently. Treat them as illustrative.

For the canonical reproduction path, use the predictions dataset at SuryanshSS1011/match-loss-to-cost-predictions, which regenerates every paper table and figure without retraining. To get a fresh checkpoint matched to a specific seed, retrain from the seed list documented in the code repo at https://github.com/SuryanshSS1011/match-loss-to-cost.

The headline experiments use the LSTM backbone, and the other architectures are loss-ablation backbones referenced in §5.C of the paper. Each .pt is a plain PyTorch state_dict loadable by the corresponding model class in src/models/ of the code repo.