OpenSimula: A Synthetic Data Generation Tool Released

OpenSimula is an open-source Python tool that implements the Simula mechanism design approach proposed in the research (TMLR) by Davidson et al., and it has been integrated into AfterImage, a dataset construction tool.

Moving away from the conventional simple 'prompt-answer' generation approach, it aims to secure controlled diversity within the reasoning space and precisely manage the axes of data variation.

Core Workflow:

• Building Factor Taxonomies: Uses an LLM to define the axes of data variation.
• Weighted Mixture Sampling: Performs sampling with weights applied based on the defined factors.
• Meta-Prompt Diversification: Applies meta-prompts to the sampled factors to increase data complexity.
• Verification and Refinement Loop: Refines the data through a Requirement Critic and a Double-critic gate for generating verifiable MCQs.

The final output generates a JSONL file containing versioned checkpoints (manifest, taxonomy bundle, sampling strategy) and approved data.