✦
Summary
Bloom is presented as an open-source, scaffolded evaluation system meticulously designed to automatically assess and surface specific behaviors within large language models. At its core, the tool operates by taking an 'evaluation configuration' or 'seed' as input. This seed is crucial as it defines the target behavior—such as political bias, sycophancy, or self-preservation—provides exemplary transcripts to illustrate the behavior, and specifies the desired types of interactions for testing. Unlike fixed evaluation methods, Bloom dynamically generates an evaluation suite that grows and adapts based on the initial seeding, making each evaluation unique and reflective of its configuration.
Key Features
Scaffolded evaluation system driven by configurable 'seed' files.
Four-stage pipeline: Understanding, Ideation, Rollout, and Judgment.
Supports flexible configuration of target behaviors and exemplary transcripts.
Intelligent batching in the Ideation Agent for enhanced generation speed.
Supports extended thinking/reasoning effort levels for model analysis.
Interactive web-based viewer for browsing generated transcripts and scores.
Unified interface for interacting with various LLM providers via LiteLLM.
Option to hide or reveal the target model's identity during evaluation.
About Bloom by Safety Research
The operational flow of Bloom is structured into four distinct, sequential pipeline stages, ensuring a rigorous and comprehensive evaluation process. It begins with the Understanding Agent, which interprets the target behavior and examples to grasp the underlying scientific motivation. Following this, the Ideation Agent creatively generates diverse evaluation scenarios designed to elicit the target behavior, utilizing intelligent batching for efficiency. The Rollout Agent then executes these generated interactions against the specified target model. Finally, the Judgment and Meta-Judgment Agents rigorously score the outcomes for the target behavior and any configured secondary qualities, with the Meta-Judgment synthesizing a comprehensive report on the findings.
The system offers substantial flexibility and control over the evaluation process, catering to various research needs, from quick local debugging to large-scale experiments managed via Weights & Biases. Users configure the entire run through a central `seed.yaml` file, specifying parameters like the target model, evaluation diversity, maximum conversation length, and whether to use extended reasoning effort or web search capabilities during scenario generation. Furthermore, Bloom supports seamless integration with external tools, including an interactive web-based viewer for browsing results and utilizes LiteLLM for unified API interaction across multiple LLM providers, facilitating model comparisons and reproducibility.
