Fix task_detector v1.2.0

v1.2.1

fix 1.2.0

1.2.0: LLM-as-a-judge Update

v1.2.0

upd 1.2.0

Release 1.1.0

1.1.0

What is new?

New Core Functions:

• Synthetic Data Generator:
  • an auxiliary module for synthetic data generation when no input dataset is provided
• Task Detector:
  • an automated task classification component for scenarios without explicit user-defined task specifications
• PromptAssistant:
  • a LLM-based component helps to interpret prompt optimization results
  • can be assigned separatly from target llm for prompting by system_model argument
  • also helps to create a synthetic dataset

Upgrades:

• Boosted HyPE:
  • New meta-prompt for optimizer HyPE that provides stronger instructive prompts
• New default llm:
  • We chose a small llm qwen3-4b-intsruct by a native huggingface launch
• New metrics:
  • Bert-score (with multilingual model) and G-Eval (experimental)

And more micro features

Check our notebook with new run examples

1.0.2

• Fixed settings of num candidates for ReflectivePrompt

Release 1.0.1

• Fixed minor bugs at the prompt evaluation stage
• Added logging of optimization steps with settings of logging completeness via verbose argument
• Expanded Readme documentation
• Fixed dependency versions in toml file

v1.0.0

First Release!

We added the main assistant interface - PromptTuner, which includes:

• Task Selection: Configuring the choice of task type: classification or generation (default)
• Metric Selection: Choosing 1 out of 3 metrics for each task type to evaluate the prompt
• Evaluation Stage: Implemented a partial prompt evaluation feature by allowing the user to submit a dataset
• LLM Customization: Selecting a custom model initialized via Langchain. By default, the built-in model configuration is called
• Two Auto-Prompting Algorithms:
  1. ReflectivePrompt – Based on evolutionary optimization methods
  2. DistillPrompt – Based on prompt distillation and Tree-of-Thoughts methods
• One Fast Adapter-Prompt: Hype - for extending the base prompt