Skip to main content
Version: v0.19.1

Release notes

Version 0.19.1 (March 13th, 2026)​

New Features​

  • [Projects] The platform now validates that experiment configurations match the selected project type, preventing misconfigured runs.

Version 0.19.0 (March 12th, 2026)​

This release includes assistant improvements along with experiment and training fixes.

New Features​

  • [Assistant] General usability and reliability improvements to the built-in assistant.

Fixes​

  • [Experiments] Fixed an issue where experiment table columns displayed ghost metrics that no longer applied to the current run.
  • [Training] Fixed a data conversion error that could cause failures in certain post-processing steps.

Version 0.18.4 (March 6th, 2026)​

New Features​

  • [Platform] Support for private Certificate Authority (CA) certificates in environments that use custom TLS infrastructure.

Version 0.18.0 (March 5th, 2026)​

This release introduces object detection as a new problem type, reworks chat templates, and aligns binary classification output with multiclass classification.

New Features​

  • [Experiments] New problem type: Object Detection, expanding beyond text-based workloads.
  • [Experiments] Chat templates redesigned for improved flexibility and consistency across model types.
  • [Experiments] Binary classification output is now consistent with multiclass classification, simplifying downstream usage.
  • [Experiments] Classification tasks now accept non-integer class labels, giving more flexibility in dataset preparation.
  • [Datasets] New built-in demo dataset: mini_textvqa_v2 for visual question answering tasks.
  • [Platform] Token access control to restrict access to platform resources.
  • [Platform] Default evaluation model upgraded from GPT 4o to GPT 5. Azure OpenAI is no longer supported as an evaluation backend.
  • [Training] Training logs are now batched, reducing noise and improving readability during long-running experiments.
  • [UI] The UI now clarifies when an evaluation uses perplexity-only metrics and skips text generation.

Fixes​

  • [Training] Fixed timezone-aware fallback handling in datetime comparisons, which could cause incorrect scheduling or display of timestamps.
  • [Training] Fixed an issue where LoRA weights were incorrectly applied during inference instead of only during training.
  • [Data Generation] Fixed a crash that occurred when a single generation failed during synthetic data generation. Individual generation errors are now handled gracefully.
  • [Experiments] Fixed the validation metrics step type for assistant tool models to use the correct numeric format.
  • [Platform] Fixed an issue that could prevent fresh installations from starting correctly.
Feedback