Repository navigation
v1.2.0
Dependencies
Refreshed renv.lock to pull the latest releases of the hubverse packages the
image wraps: hubPredEvalsData 1.1.1 -> 1.2.0, hubEvals 0.2.0 -> 0.3.1,
hubUtils 1.2.0 -> 1.2.1, and hubData 2.2.1 -> 2.2.2 (#47).
The user-visible changes these bring to the generated dashboard data:
predevals-options.jsontargets now carrytarget_nameandtarget_units,
read from the hub's tasks config. The dashboard usestarget_nameto label
target menu items (hubverse-org/hubPredEvalsData#21).scores.csvrows are now sorted on the(model_id, by)key before writing,
so the file is byte-stable across runs. Output can be diffed or
md5-compared to confirm a no-change situation
(hubverse-org/hubPredEvalsData#25).- Models scored only on a later output type are no longer silently dropped
fromscores.csv. The per-output-type score frames are now merged with a
full join rather than a left join anchored on the first output type, so
which models survived no longer depends on the order metrics happen to be
listed in the config (hubverse-org/hubPredEvalsData#75). - Relative-skill scoring no longer aborts the run on single-model input, or
on disaggregated comparison groups that cannot be compared. Such groups are
reported as relative skill1orNAwith a warning, and their absolute
scores are returned unchanged (hubverse-org/hubEvals#75,
hubverse-org/hubEvals#135). - Scoring on transformed scales with
transform_appendnow returns one row
per scale instead of silently averaging across scales
(hubverse-org/hubEvals#122). - Oracle output is now read correctly from cloud (S3) hubs whose
admin.json
declares a non-parquet submission format such ascsv. The cloud sync always
writesmodel-output/to parquet on S3 regardless of the declared format, so
such hubs previously came back as an empty connection
(hubverse-org/hubData#148).
Full Changelog: v1.1.1...v1.2.0