HCAI-Lab/w2-consensus-deepdive-unlearning-artifacts / social-data-attribution-w2 /src /data_attribution /evaluation /README.md
Evaluation Package Scope
OlmoBaseEval
I keep OlmoBaseEval support as a separate expansion track from the current SocialTDA paper benchmark set.
The code in this package may live with the shared evaluation infrastructure
because it reuses OLMES export, query-manifest conversion, and TrackStar
analysis paths. That does not make base_olmobaseeval result artifacts part of
the SocialTDA paper outputs.
Current merge scope:
- Registry and conversion support for cloze-compatible OlmoBaseEval probes.
- Tests that prevent accidental export or aggregation of unintended probes.
- No tracked cluster manifests, rendered result pages, or paper-facing OlmoBaseEval result tables.
Result policy:
- Treat prior
base_olmobaseevalPACE and WIP outputs as separate-analysis artifacts. - Do not mix OlmoBaseEval results into SocialTDA paper result folders, report tables, figures, or claims unless a later release explicitly promotes that scope.
- Keep cluster-specific paths, run manifests, and large generated outputs out of the implementation PR unless they are reviewed as a separate results PR.
Xet Storage Details
- Size:
- 1.11 kB
- Xet hash:
- f8e08f9bb0e9b61fd6ad8e67ad1d78c33d6f43fe7a550151a65841f8edad6baf
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.