Bloom Reviews — Discover what people think of this product.

Bloom is an open-source framework for automated behavior evaluation of large language models, designed for safety researchers and ML engineers who need reproducible, configurable probing of complex model behaviors. Instead of relying on fixed prompt sets, Bloom grows an evaluation suite from a seed configuration that defines a target behavior, example transcripts, and interaction parameters.

Using a simple seed.yaml and behaviors.json, you can specify behaviors such as sycophancy, political bias, or self-preservation, then automatically generate rich interaction scenarios against models like claude-sonnet-4. Bloom orchestrates multi-turn conversations, logs every pipeline stage to structured JSON, and stores transcripts for further analysis.

Key capabilities include:

• Scaffolded behavior exploration driven by seed configurations
• Configurable diversity, modality, and interaction depth for nuanced probing
• Support for multiple LLM providers via API keys in .env
• Interactive transcript viewer with scoring, justifications, and filtering

By pairing seeds with version-controlled configurations, Bloom enables reproducible safety evaluations, systematic sweeps across behaviors, and rigorous comparison of model variants under evolving behavioral test suites.