Canonicalized Jupyter notebooks are structured documents that can mix source code, markdown, tracebacks, HTML, SVG, metadata, and large binary outputs encoded inside JSON. This task treats the corpus as a domain-specific compression problem rather than a generic text or blob compression problem.
The agent receives a visible corpus of canonicalized Jupyter notebooks and must build a lossless compression pipeline for them. The submission is not a single executable with one verb. It is a three-stage system: fit a model on the visible data, compress a hidden holdout, and decompress it back to the original tree exactly.
fit can build learned artifacts or dictionaries from visible data.compress must minimize hidden holdout size.decompress must reconstruct every file byte-for-byte with the same relative paths.The verifier canonicalizes notebooks, runs the submitted fit/compress/decompress pipeline, enforces a strict lossless round trip, and then computes geometric-mean compression ratio across the hidden holdout. The reported score is reduction (1 - r), so higher is better.
This is a large CPU-and-disk task: 16 vCPU, 32 GiB RAM, 150 GiB scratch space, no GPU, and no internet access. The hidden holdout is large enough that aggressive but invalid tricks are easy for the verifier to catch during reconstruction.