Tags: DaoyuanLi2816/mini-verl
Tags
miniVERL v0.6.3 - security, artifact integrity and trust boundaries bridge doctor treats an exported bundle as hostile input. It no longer imports the reward scaffold; the file is parsed with ast.parse and every definition-time position is audited, including class bases, metaclass and other class keywords, annotations and type-parameter bounds. Adapter safetensors are validated past the header through dtype and shape byte arithmetic, offset ordering, contiguity and full payload coverage. What a bundle declares about itself is reported separately from what this process recomputed, because a bundle's own SHA256SUMS can prove internal consistency only. An input file can never also be an output file. Malformed extension sidecars fail the conversion instead of being read as empty. Dataset conversion streams row groups instead of materializing the whole table. Portable metadata privacy detects credentials, not only absolute paths, and never reports matched values. Stable and development version claims have one canonical source. Every public scientific figure is readable at 390px with no exemption list. Tracked text is gated for mojibake and byte-order marks. No new scientific result, benchmark or distributed-execution claim. Every frozen scientific artifact remains byte-identical.
miniVERL v0.6.2 Bridge correctness and responsive visual hardening. Unresolved Hydra/OmegaConf interpolation can no longer reach a runnable recipe. Bridge outputs are stem-specific and published transactionally behind an exclusive per-stem reservation, with --overwrite required to replace an existing family. bridge doctor reports explicit tokenizer verification levels and separates portable-metadata, dataset-content and model-weight privacy, where not_inspected never means passed. The Alignment Lab figures are readable at 390 px through real vertical layouts, metric coverage became an accessible responsive table, and the browser visual gate now measures actual rendered geometry. Every frozen scientific artifact remains byte-identical. No distributed verl execution and no OPD/PPO semantic parity is claimed.
PreviousNext