Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
ICML 2026 paper 3092 reproduction
Independent, budget-capped reproduction of “Decomposing
Out-of-Distribution Error in Conditional Flow Matching via Wasserstein
Geometry” (OpenReview orzcDIBxa0).
The official implementation was not discoverable on GitHub as of 2026-07-24. This repository therefore contains:
repro_decomposition.py: an exact numerical audit of the three-term Wasserstein decomposition for equal-variance one-dimensional Gaussians.repro_diagnostics.py: a synthetic two-moons conditional flow-matching experiment that tracks interpolation sparsity, a Jacobian-excess geometric distortion proxy, in-distribution fit, and OOD error through training.mnist_cfm_wandb.py: an image-scale MNIST experiment holding out digit 8, comparing baseline CFM with the paper's targeted Mixup intervention, and logging latent-Wasserstein diagnostics plus samples to Weights & Biases.run_mnist_canary.sh/run_mnist_campaign.sh: pinned cloud entrypoints for the short W&B/GPU validation and the 3-seed campaign.submit_mnist_job.py: credential-safe HF Jobs submitter with a 20-minute canary cap and a 5-hour ($4 at current L4 pricing) campaign cap.make_figures.py: generates the interactive figures used in the Trackio logbook from the raw JSON results.
Primary sources:
- Paper page: https://openreview.net/forum?id=orzcDIBxa0
- ICML event: https://icml.cc/virtual/2026/poster/61565
- Official poster: https://icml.cc/media/PosterPDFs/ICML%202026/61565.png
- MNIST dataset: https://huggingface.co/datasets/ylecun/mnist
- Downloads last month
- 77