ru1ch3n/PDE-OBS-excluded-checkpoints

PDE-OBS: checkpoints of the runs excluded from the October 2026 results

Weights-only checkpoints of the training runs of the PDE-OBS October 2026 campaign that are not part of the published results. The published models and results are in the companion repositories ru1ch3n/PDE-OBS (models) and ru1ch3n/PDE-OBS (data and results); the benchmark code is at https://github.com/ru1ch3n/PDE-OBS. No checkpoint in this repository contributes a score to the result tables of the paper (the failed attempts appear there only as failure rows); the runs are archived so that the campaign's record is complete.

manifest.json lists every run with its exclusion class and the SHA-256 and size of every file. The classes:

• development_screen: recipe and learning-rate screens (V2dev, V4dev, H0dev, H3dev, H0prof); the frozen recipes chosen from them are in configs/paper/october/recipes.yaml of the benchmark code.
• superseded_plan_block: the earlier plan stages H2, H3, H5, H7 and P2, replaced by the V2 plan of 2026-10-05.
• failed_training: attempts that ended in a training pathology under a frozen recipe (constant-output collapse or a sustained large gradient; the result tables list them with their failure).
• voided_attempt: attempts repeated under the same run id after an infrastructure failure (the directory name carries the .void-<utc> suffix of the campaign). The four such directories hold truncated checkpoint files (the attempt was voided while the file was being written); they are the runs listed under not_included and the class has no entry.
• pipeline_check: the PIPECHECK runs that compared the frozen and the optimised training loop.
• pre_campaign_experiment: the e2 DeepONet runs of 2026-10-01/02.

Layout

<cohort>/<run_id>/
    weights.pt               {"format": "pdeobs-october2026-weights/1", "run_id": ..., "model_state": <state dict>}
    weights_dev_ep<N>.pt     intermediate development checkpoints, where the run kept them
    model.json               cohort, exclusion class, queue ledger status, de-identified failure record, numeric progress,
                             per checkpoint: original and released SHA-256 and sizes, tensor counts, dropped keys
    history.json             per-epoch training history (numeric fields only), where the checkpoint held one
manifest.json                every run, class, file, SHA-256 and size; counts by class and cohort

torch.load(path, map_location="cpu", weights_only=True)["model_state"] loads a weights file. Optimizer, scheduler and RNG states and the resolved run configuration (which held machine paths) are not included. Strings in the failure records that named a filesystem path, account, node, job or site were replaced by the placeholders <PATH>, <USER>, <NODE>, <JOB> and <SITE>; in the eight PIPECHECK run ids the account name was replaced by a letter. The run ids are otherwise those of the campaign.

Runs not included

4 of the 452 excluded run directories could not be archived; manifest.json lists them under not_included with the reason: no checkpoint could be converted (4).

Citation

Ruichen Xu, Siyao Wang, Fang Wan, Jiacheng Qiu, Wenhan Gao, Jiaxing Zhang, Linsey Pang, Ravid Shwartz-Ziv, Yann LeCun and Yuefan Deng. PDE-OBS: Controlled Evaluation Across Observation Patterns. arXiv preprint arXiv:2609.36521, 2026.