Zentrix Agentic Workbench
@deslito
A short summary of what the server does. Example: Modular agentic AI workflows for dataset cleaning and orchestration.
概要
What is Zentrix Agentic Workbench?
Zentrix Agentic Workbench is a comprehensive pipeline for generating a high-quality, balanced dataset for training and evaluating conversational AI models with multi-tool and file-access capabilities. It combines synthetic data generation with the integration and normalization of external datasets from Hugging Face.
How to use Zentrix Agentic Workbench?
Run the full pipeline with python main.py to generate everything from scratch (synthetic data and external downloads) or use python main.py --merge-only to quickly create the training set from existing processed datasets. Additional options include --force, --skip-downloads, and --num-samples. Evaluation is done via evaluate_with_mcp.py with configurable MCP endpoint, API key, and sample limit.
Key features of Zentrix Agentic Workbench
- Combines synthetic data with external Hugging Face datasets
- Generates multi-tool invocation and file-access dialogues
- Validates and normalizes dataset schemas automatically
- Outputs a final balanced training set as
training_set.jsonl - Supports evaluation using MCP.so for quality metrics
- Offers quick merge-only mode for iterative development
Use cases of Zentrix Agentic Workbench
- Training conversational AI models that invoke multiple software tools
- Building file-access capable dialogue agents from diverse data
- Evaluating dataset quality with exact match and tool invocation accuracy
FAQ from Zentrix Agentic Workbench
What does the final output look like?
The pipeline produces final_datasets/training_set.jsonl, a shuffled, balanced dataset containing 100% of records from both multi-tool and file-access tasks.
How do I skip external dataset downloads?
Use the --skip-downloads flag when running main.py to generate only synthetic data.
Where are the downloaded external datasets stored?
External datasets are downloaded from Hugging Face into the downloaded_datasets/ directory during a full pipeline run.
What metrics does the evaluation track?
The evaluate_with_mcp.py script tracks exact match accuracy, response structure consistency, tool invocation accuracy, and common failure patterns.
What runtimes or dependencies are required?
Python 3 and the packages listed in the project files. The pipeline uses MCP.so for evaluation and requires a valid API key and endpoint URL.
