myx160/rl-post-train-checkpoints

RL_post_train archival backup (2026-10-04)

Public archival backup, preserving original project-relative paths. Not a newly validated benchmark model release. The Before Task 0 result is 50/250 = 20.0%; a formal trained RECAP After result has not yet been established. CFG step-20 checkpoints are wiring/preflight artifacts, not evidence of improvement.

The upstream RECAP data is RLinf/RECAP-Libero10-Task0-48succ-Data at revision 75b382d2c066bcedd8b030285b45c856913f1497. The SFT starting checkpoint is RLinf/RLinf-Pi05-LIBERO-SFT at revision 45ccfcc4e28634f1576ebf78cab0fbe2fd82432d. Gemma-derived files are subject to the included Gemma Terms, Section 3.2 use restrictions, and Prohibited Use Policy; these terms must be honored by all recipients. Project-trained Value/CFG files are modified derivatives, and random/smoke files are not formal models. See LICENSES and the preserved experiment documents.

Collection: https://huggingface.co/collections/myx160/rl-post-train-6ac23cee2497d9f46d613e3c

Companion repositories: models, datasets and artifacts. Runtime-cache tar files contain ordinary files only, with project-relative member names and explicit contents manifests. Credentials, Git metadata, Unix sockets, lockfiles and Python bytecode are excluded and remain local.

Restore files using huggingface_hub.snapshot_download with repo_type='model' or 'dataset' and a revision from the local verification receipt. Extract cache tar archives only into the matching project workspace, after inspecting their contents manifests.