Learning-Order GFlowNets — private research handoff
Start with AGENT_REPRODUCTION.md. It describes how to recreate all 15 paper figures, tables, and the PDF, and how to reevaluate saved models.
| Archive | Contents |
|---|---|
| gfnx_paper_reproduction_20261030.zip | Code, paper, exact compact figure inputs; CPU plotting, about13MB |
| gfnx_results_20261030.zip | Historical structured results, about3.76GB |
| gfnx_weights_20261030.zip | Retained paper models and all14332 staged SNR checkpoints, about77.5GB |
| gfnx_weights_legacy_20261030.zip | Complete Orbax model directories and old Sudoku checkpoint tuples, about1.54GB |
Download each needed ZIP and its matching .zip.sha256 sidecar; verify with
sha256sum -c NAME.zip.sha256. Start with the compact archive to reproduce figures.
It includes a compiled reference PDF. Extract the two weights ZIPs into separate
folders to preserve their separate weights_coverage.json reports.
All requested SNR snapshots were exported. Not every older quality-curve point had its own retained weights. Existing curves remain reproducible from archived metrics. The full results ZIP records486 unreadable historical source entries; exported weights had no read errors. See the agent guide and coverage reports before fresh evaluations. This repository is private. No full Sudoku training corpus, compiler cache, or complete optimizer/replay training-resume archive is included. Original sources are unchanged.
Code: https://github.com/d-tiapkin/gfnx-private/tree/sudoku_and_snr Packaged revision: 2a9eb3a7078fa98bab548eb68684e6c4267a0191.