CSP research checkpoints
Public training checkpoints and scalar learning curves for ten CSP research runs (Push-T, Reacher, Cube and Two-Room). Experimental models, not a claim of benchmark success or production readiness. One training seed per configuration.
Each runs/<run-id>/config.json records the scientific settings. latest.json
identifies the current and previous complete resume states and their SHA256s.
Resume files include neural modules, optimizers, SIGReg buffers, RNG states,
module modes and training position. They intentionally omit original filesystem
paths and credentials. A transport adapter restores the locally pinned config
and source identity before invoking the unchanged native trainer.
Weights-only milestone exports and scalar training logs are also retained. No datasets, images, credentials, environment dumps, private communications, or unrelated project artifacts are included. Original cluster checkpoints are preserved. Download only checkpoints from sources you trust: full PyTorch training-state files use pickle and are not safe to load from unknown authors.
Uploads are transactional and verified by downloading the pinned commit and checking hash and state equality. Node-local scratch is disposable; restarts restore the latest confirmed remote checkpoint, never assume scratch persists. Checkpoint publication can lag the currently running update. Abrupt preemption can lose the updates since the latest completed upload, not the whole run.
Two recovery slots are retained in the current tree per run. The owner explicitly
authorized periodic old-history compaction for THIS dedicated repository on
2026-09-24. All current files, including both slots and all milestone weights,
remain; older replaced checkpoint versions are not permanent archives.
Do not use mutable main as an
unversioned scientific result: record exact checkpoint hashes for evaluations.
Transport schema: m2-csp3-hf-20260924-public-v1.