mpstoryfans/llapdiff-checkpoints

LLapDiff Checkpoints

This repository hosts checkpoint archives for the LLapDiffusion project.

Archives

• llapdiffusion_baseline_checkpoints.zip: longest/default-horizon extrapolation baseline checkpoints. It contains 56 checkpoints across 7 datasets and 8 baselines.
• llapdiffusion_longest_horizon_artifacts.zip: LLapDiff longest-horizon artifacts trained from commit 59f4427b3a0c71e941f6ff4f7f047f80cad2dfd6.

Baseline archive

The baseline archive has two completion groups:

• 35 full-loop checkpoints for dlinear, mtan, patchtst, t_patchgnn, and timegrad.
• 21 very heavy checkpoints for neuralcde, mr-diff, and contiformer.

The one-batch checkpoints are included for completeness and load/evaluation plumbing checks; they are not comparable full-training baseline results. Checkpoint names follow checkpoints/<dataset>_h<horizon>_<method>.pt. MANIFEST.csv records file sizes, SHA256 checksums, and completion_mode.

Current llapdiffusion_baseline_checkpoints.zip SHA256: 3afc405489dd061a92bb5b66d82cca4c12ae56b43fefc12e6a191ded9d22d3f4

LLapDiff artifact archive

The LLapDiff artifact archive covers bms_air h168, uci_air h168, physionet h12, noaa_us h168, noaa_uk h168, us_equity h100, and crypto h100. For each dataset, it includes VAE ELBO/reconstruction best checkpoints, summarizer best, and LLapDiff best.

Historical Hugging Face revisions may contain earlier archive states.

CSDI 30% Imputation Checkpoints

llapdiffusion_csdi_mask30_imputation_checkpoints.zip contains CSDI target-horizon imputation checkpoints for PhysioNet h12, Crypto h100, and NOAA UK h168. These were trained/evaluated with a random 30% observed-target holdout and 25 evaluation samples. The archive is separate from the extrapolation baseline archive and the LLapDiff artifact archive.