بلوم: تقييم سلوك آلي مفتوح المصدر لنماذج اللغات الكبيرة

Bloom is an open-source tool designed to automate the evaluation of behaviors exhibited by Large Language Models (LLMs). It provides a scaffolded system that takes an evaluation configuration (a “seed”) as input, defining the target behavior, exemplary transcripts, and desired interaction types. This input drives the generation of a comprehensive evaluation suite that attempts to uncover the specified behavior. This approach allows for a dynamic evaluation suite tailored to specific behaviors, unlike fixed elicitation techniques. Here’s a breakdown of Bloom and its capabilities:

Key Features and Functionality

• Behavior-Driven Evaluation: Bloom allows users to specify the behaviors they want to evaluate, such as sycophancy, political bias, or self-preservation. It then generates interactions designed to elicit these behaviors.

• Seed Configuration: The evaluation process is driven by a “seed” configuration file (seed.yaml) that defines the target behavior, examples, interaction parameters, and model configurations. This ensures reproducibility, as Bloom evaluations should always be cited along with their full seed configuration.

• Modular Pipeline: Bloom’s pipeline is divided into four distinct stages:

  • Understanding Stage: Analyzes the target behavior and provided examples to understand the underlying mechanisms and motivations.
  • Ideation Stage: Generates diverse base evaluation scenarios and their variations. This stage is critical for creating a comprehensive and varied test suite.
  • Rollout Stage: Executes the evaluation scenarios with the target model, simulating conversations or environments.
  • Judgment Stage: Scores each rollout for the presence of the target behavior and other configurable qualities. This stage also includes a meta-judgment that analyzes the overall evaluation suite.

• Interactive Viewer: The tool provides an interactive viewer to browse transcripts, view conversation flows, see judgment scores, and filter/search through transcripts. This facilitates detailed analysis of the evaluation results.

• LiteLLM Integration: Bloom utilizes LiteLLM to handle API calls to various LLM providers, offering a unified interface to services like OpenAI, Anthropic, OpenRouter, and Amazon Bedrock.

• Weights & Biases Integration: Bloom offers Weights & Biases (wandb) integration for large-scale experiments and tracking. This allows for better experiment management and analysis.

Getting Started with Bloom

To start using Bloom, follow these steps:

• Install Dependencies: Clone the repository and install the necessary Python packages using uv pip install -r requirements.txt.
• Configure API Keys: Add API keys for the LLM providers you want to use (e.g., OpenAI, Anthropic) to a .env file.
• Define Target Behavior: Define the behavior you want to evaluate in behaviors/behaviors.json.
• Provide Examples (Optional): Add example transcripts of the target behavior to the behaviors/examples/{your-behavior-name}/ directory.
• Configure seed.yaml: Customize the seed.yaml file, setting parameters such as the target behavior, examples, number of evaluations, target model, and diversity.
• Run the Pipeline: Execute the pipeline locally using python bloom.py --debug.

Advanced Features

• Intelligent Ideation Batching: The ideation stage uses intelligent batching to generate multiple scenarios per API call, improving efficiency.

• Resuming Sweeps: Bloom allows you to resume interrupted Weights & Biases sweeps, which is helpful for large-scale experiments that may encounter API failures or timeouts.

• Extended Thinking/Reasoning Effort: Bloom supports extended thinking for Claude and OpenAI models, allowing for more detailed reasoning during the evaluation process.

• Evaluating Multiple Target Models: Bloom supports evaluating multiple target models with identical evaluation scenarios.

Conclusion

Bloom is a valuable tool for researchers and developers looking to evaluate the behavior of LLMs in a structured and automated way. Its modular pipeline, interactive viewer, and integration with LiteLLM and Weights & Biases make it a powerful platform for understanding and mitigating potential risks associated with LLMs. By providing a configurable and reproducible evaluation framework, Bloom contributes to building safer and more reliable AI systems.

Source: safety-research/bloom