Skip to main content

3.20 Release Notes

3.20.4​

Improvements & Enhancements​

  • Optimized and accelerated the synthetic data generation process for custom content policies
  • Enabled GPU slicing for input guardrails running in EKS

Bug and Vulnerability Fixes​

  • Fixed bugs in custom output guardrail data generation
  • Resolved inconsistencies in policy training status and header status
  • Temporarily removed the cancel training feature. It will be re-enabled once the internal logic is fixed

3.20.3​

Improvements & Enhancements​

  • Added new Dynamo AI logo to the UI
  • Added minor upgrades to the UI policy creation workflow
  • Added support for using custom LLM models during data generation in policy creation
  • Added tracing for DynamoGuard inference

Bug and Vulnerability Fixes​

  • Fixed policy creation UI bugs
  • Fixed finetuning job with new base image library
  • Resolved keycloak running issues on HTTP servers

3.20.2​

Bug and Vulnerability Fixes​

  • Added CUDA library to model images to support GKE nodes

3.20.1​

Improvements & Enhancements​

  • Enabled inference API timeout to be configurable
  • Improved DynamoGuard RAG hallucination models

Bug and Vulnerability Fixes​

  • Fixed bugs in DynamoEval progress bar and deep dive UI
  • Updated Output guardrail thresholding to follow same scheme as input guardrails
  • Updated the name of the Hallucination Guardrail submetrics.

3.20.0​

New Features​

  • [Beta] View and update policy definitions while reviewing examples
  • [Beta] Keyword blocklist policy: Block prompts and responses containing specific words or phrases
  • [Beta] New Platform Admin user role for billing and platform management
  • [Beta] New controller to enable dynamic model deployments in the cluster for input custom content policies
  • [Alpha] Connect custom remote models as AI Systems to run DynamoEval tests on
  • [Alpha] DynamoEval progress bar preview for PII extraction and bias/toxicity tests
  • [Alpha] Connect custom models for data generation during the DynamoGuard policy creation process
  • [Alpha] Duplicate default and custom content policies

Improvements & Enhancements​

  • Improved input guardrail model training process, resulting in more performant guardrails
  • Improved output safety model