Studying how models train often requires saving sequences of checkpoints — checkpoint trajectories — but the storage cost of these snapshots limits the temporal resolution at which training can be examined. A new paper on arXiv introduces ULPack, a lossless compression method designed for exactly this setting.
ULPack is reference-conditioned: it compresses each checkpoint relative to a reference checkpoint, exploiting the fact that adjacent checkpoints are numerically close. The paper notes that existing lossless codecs either operate on XOR bit patterns or otherwise fail to take advantage of this closeness. By conditioning on a reference, ULPack aims to remove the redundancy between neighboring snapshots without any loss.
Because the compression is lossless, the reconstructed checkpoints are identical to the originals, so the method does not trade fidelity for storage savings. The abstract does not report specific compression ratios or benchmark numbers, so the magnitude of the savings is not stated in the source.