Repository navigation
EVA v1.2.1 repairs the existing dense-control training and checkpoint-loading entries and includes the recovered author source and configuration evidence.
The release adds public checkpoint/dataset SHA256 inventories and manuscript-to-artifact mappings, and identifies the original ProGen3 scoring CLI and scorer at a fixed upstream commit. Benchmark inputs, raw results and Hugging Face revisions are unchanged.
Validation: Python 3.10 and 3.11 CPU CI passed for this exact commit, including regression tests, documentation/citation checks, source-archive execution and wheel installation outside the checkout. The dense implementation also has a separate successful full-runtime parameter-update and exact save/reload check documented in docs/DENSE_CONTROL.md.
- Source commit:
f3447eb4f0e0ac061c283d2ea1e00076ddef7ed7 - CI: https://github.com/GENTEL-lab/EVA/actions/runs/34410621031
- Code archive DOI: https://doi.org/10.6084/m9.figshare.33490096.v4
- Source ZIP SHA256:
8a291df4da14b20faab56081967dd33a3196b254bf1398d4b2159abd1652134a
The source ZIP contains all 776 tracked files. The metadata ZIP contains a per-file SHA256 manifest. v1.2.0 remains unchanged.