NVIDIA/cosmos-curator v2.2.0
NVIDIA/cosmos-curator
Captured source
source ↗published Jun 11, 2026seen 1dcaptured 1dhttp 200method plain
Release v2.2.0
Repository: NVIDIA/cosmos-curator
Tag: v2.2.0
Published: 2026-06-11T18:50:45Z
Prerelease: no
Release notes:
Breaking Changes
- Remove legacy AV pipeline and PostgreSQL helper modules; see
Removedfor affected paths. - Remove the
unifiedPixi environment; usedefaultfor stages and commands that previously
selected unified.
Added
- Add
cosmos-curator pipelineconfig tooling for Ray Data pipeline configs: templateprints editable config templates.validatechecks resolved JSON/YAML configs.renderemits the canonical resolved config used for execution.schemaprints the supported config JSON Schema.presets listandpresets showinspect packaged presets.- Add schema-validated Ray Data
video_splitconfigs with defaults, packaged presets, template
profiles, strict field validation, and --set dotted-path overrides.
- Add split-comparison measurement schema and field-spec catalog for typed measure/evaluate flows.
- Add caption quality statistics to split benchmark summary metrics.
- Add shared caption embedding helpers for split-comparison measurement code.
Fixed
- Reject invalid curator-side
uint8video preprocessing configurations before execution. - Force model-side preprocessing for Nemotron in the split pipeline.
- Centralize vLLM preprocess-mode handling across supported vLLM model wrappers.
- Return NVCF API response status codes when surfacing API errors.
- Improve image-build CI behavior and allow pre-commit setup to run on macOS.
Changed
- Split Pixi tooling into narrower developer, tooling, cluster, runtime, and model-specific
environments to reduce development-environment weight.
- Keep the Pixi developer environment slimmer by moving runtime-heavy dependencies out of the
general dev feature.
- Reduce Slurm end-to-end CI inode usage.
- Update Ray Data split pipeline internals to execute through resolved config objects.
Removed
- Remove the legacy AV pipeline package:
cosmos_curator.pipelines.avis no longer included.- The old AV pipeline reference documentation is removed.
- Remove PostgreSQL helper modules that supported the legacy AV pipeline:
cosmos_curator.core.managers.postgres_cliis no longer included.cosmos_curator.core.utils.dbis no longer included.- Remove the
unifiedPixi environment after folding its common GPU-capable runtime dependencies
into default.
Documentation
- Add pipeline-config design documentation for schema-versioned config files and config tooling.
- Add Pixi environment design documentation for the split developer/runtime environment layout.
- Refresh deprecation and split-comparison design notes for the new config-driven and measurement
workflows.
Notability
notability 3.0/10routine version update of a data curation tool