Skip to main content

Snorkel Flow v25.4 (LTS) release notes

Breaking changes​

  • Legacy notebook functionality on Kubernetes has been removed.

Features and improvements​

Annotation​

  • Improved role-based UI restrictions: Labelers (Annotators) now see only tabs relevant to their workflow.
  • You can now perform bulk labeling for sequence tagging multi-schema annotation (Text).
  • Label schemas can now be included during annotation task creation.
  • The loading state for annotations has been streamlined for better performance.
  • The delimiter for highlight UIDs has been updated.
  • Added Data Explorer tab to datasets page where users view individual data points inside a dataset.

Prompt development​

  • You can now run your prompt on just the first N rows of data for faster testing and iteration in both traditional and LLM-as-a-judge (LLMAJ) prompt development.
  • The prompt editor now correctly highlights columns regardless of case sensitivity.
  • Ground truth filter errors in LLMAJ and prompt development have been fixed.

ML tasks: text​

  • Error reporting for Ground Truth upload has been improved, with partial writes now disabled.

Infrastructure​

  • Added SDK function sf.get_model_provider_status() to validate operational status of foundation model providers.
  • Added support for running concurrent LLM inference requests on Bedrock.
  • Improved responsiveness of helm charts.

User interface​

  • Dataset splits now appear in the UI immediately upon first upload.

Integrations​

  • You can now use Llama API prompt inference requests via the Custom Inference Service.

SDK​

  • Fixed a circular dependency issue when installing a wheel.

Bug fixes​

ML tasks​

  • Fixed an issue with Ground Truth uploads.

SDK​

  • The workspace_uid parameter has been removed from set_secret.

Known Issues​

Data upload​

  • New data sources do not have embeddings generated if that feature is not activated.
  • Uploading large CSV files can show unrelated errors during data upload.
  • There may be dataset size discrepancies between the actual file size and what is shown in the GUI.
  • Downloading PDFs with the https URL fails.

Annotation​

  • For PDFs, the annotation filter for negative ground truth doesn't work.
  • In Annotation Trace view, the first document might not load existing annotations on page load.
  • Focus on free text multi-schema annotation inputs may act erratically, especially when using Tab to switch between fields.

Data development​

  • You may receive an error that certain datapoints are not in index after resampling a dev split.
  • Studio /dataset and /advanced-lf-state endpoints error out with a cryptic error message when there is no span.
  • Drop context_pages and page_docs need to be dropped at /dataset & /context-dataset.

Prompt development​

  • You may encounter errors when creating an LLMAJ with a previously deleted exact name.
  • Errors may occur when loading ground truth for LLM responses in prompt development.
  • Inaccurate counts may appear in traces batch.
  • Trace step pagination may not load all steps of a trace when loading a long trace.

Evaluation​

  • Multi-trace dictionary values for trace steps have limitations with int, float, and bool types.
  • Benchmark graphs may load before the run is completed.
  • LLMAJ Schema filter is not properly scoped to the selected benchmark.
  • GT filter in All Prompt views should be 'present/absent', not asking for the value.
  • In Prompt table view, truncated text may be difficult to read.
  • You may experience double-loading when clicking to the next page in Traces.

SDK​

  • The SDK aggregate_annotations() method may fail.