SyGra 2.0.0 Studio Release

SyGra 2.0.0 introduces Studio, an interactive visual environment designed to transform synthetic data generation from a manual process of editing YAML files and using terminals into a transparent, visual craft. Studio allows users to compose flows on a canvas, preview datasets, and tune prompts with inline variable hints while maintaining full compatibility with the underlying SyGra platform by generating corresponding graph configurations and task executor scripts.

Visual Workflow Design and Composition

SyGra Studio enables the creation of synthetic data pipelines through a drag-and-drop interface. Users can build complex logic by utilizing various node types and configuration options:

  • LLM Nodes: Users can select configured models (such as gpt-4o-mini), write prompts, and define output variables (e.g., storing a result in story_body).
  • Advanced Logic: The environment supports structured outputs, tool attachments, Lambda nodes, and Subgraph nodes for reusable logic or branching behavior.
  • Dynamic Prompting: Typing { inside a prompt editor surfaces all available state variables instantly, allowing for seamless referencing of data from previous nodes.
  • Multi-LLM Settings: The detail panel provides access to model parameters and settings for parallel generations.

Data Source Integration and Validation

Studio streamlines the initial phase of synthetic data generation by providing guided forms for data source configuration and validation:

  • Supported Connectors: Users can connect to Hugging Face, the local file system, or ServiceNow data sources.
  • Immediate Preview: After entering parameters such as repo_id, split, or file path, users can preview sample rows to ensure the data is correct before execution.
  • Automatic Variable Mapping: Column names from the data source are automatically converted into state variables (e.g., {prompt}, {genre}), eliminating manual wiring and guesswork.

Execution, Observability, and Debugging

SyGra Studio provides real-time monitoring and debugging tools to ensure the quality and cost of synthetic data runs:

  • Real-time Streaming: The Execution panel streams node status, token usage, latency, and cost in real time during a workflow run.
  • Observability: Detailed logs and execution history are stored in .executions/runs/*.json, allowing users to compare results against prior executions.
  • Debugging Tools: The interface includes inline logs, breakpoints, and Monaco-backed code editors for rapid iteration.
  • Run Configuration: The Run modal allows users to adjust batch sizes, record counts, retry behavior, and temperatures without modifying the underlying YAML configuration.

Model Support and Integration

Studio supports a wide array of model providers and custom endpoints to ensure flexibility in generation:

  • Guided Configuration: Users can validate models from OpenAI, Azure OpenAI, Ollama, Vertex, Bedrock, vLLM, and custom endpoints.
  • Infrastructure Compatibility: Because Studio generates the same YAML/JSON artifacts written to tasks/examples/, the visual designs are fully compatible with the same execution engine as the CLI.

Example Use Case: Glaive Code Assistant

SyGra Studio can execute existing workflows, such as the Glaive Code Assistant. This specific workflow ingests the glaiveai/glaive-code-assistant-v2 dataset and employs a loop of two LLM nodes—generate_answer and critique_answer—linked by a conditional edge. The process continues to loop until the critique node returns "NO MORE FEEDBACK," resulting in synthetic data ready for model training or evaluation.

Sources