Hypogenic

This automates hypothesis generation and testing on tabular data using LLMs, which is genuinely useful if you're doing empirical research and tired of manually formulating hypotheses. It offers three approaches: pure data-driven generation, literature-plus-data integration, or mechanical combination of both. The framework handles the full loop from generating 10-20 testable hypotheses to running inference and iterative refinement based on validation performance. Setup requires following HuggingFace dataset conventions and writing YAML configs with prompt templates, which adds some overhead but gives you control. The benchmarks claim 9-16% improvements over baselines and 80%+ hypothesis diversity. Redis caching is smart for cutting API costs during iteration. Best suited for research domains like deception detection or content analysis where you have structured data and want systematic exploration rather than ad-hoc prompting.