Bloom: أداة مفتوحة المصدر للتقييم الآلي لسلوك نماذج اللغة الكبيرة

Bloom is an open-source tool designed for the automated evaluation of Large Language Model (LLM) behaviors. It provides a structured framework for assessing specific behaviors, such as sycophancy or bias, through the generation of customized evaluation suites. This differs from static evaluation methods by adapting the evaluation based on an initial “seed” configuration, making reproducibility dependent on the seed. This blog post summarizes Bloom’s key features, functionalities, and usage instructions, providing a comprehensive overview for developers and researchers interested in leveraging it for LLM evaluation.

Key Features of Bloom

• Automated Evaluation Suite Generation: Bloom automatically creates evaluation suites based on a user-defined “seed” configuration. This seed includes target behaviors, example transcripts, and desired interaction types.
• Customizable Evaluation: Bloom is highly customizable. Users can define target behaviors, provide example transcripts, and configure various parameters to tailor the evaluation process.
• Modular Pipeline: The evaluation process is divided into four distinct stages: Understanding, Ideation, Rollout, and Judgment. This modularity allows for targeted debugging and iterative refinement of the evaluation process.
• Integration with Multiple LLM Providers: Bloom supports multiple model providers through LiteLLM, including OpenAI, Anthropic, OpenRouter, and Amazon Bedrock, providing flexibility in selecting the target model for evaluation.
• Interactive Transcript Viewer: Bloom includes an interactive viewer for browsing evaluation transcripts, facilitating analysis of conversation flows, judgment scores, and justifications.
• Weights & Biases Integration: Bloom integrates with Weights & Biases for large-scale experiments, enabling efficient tracking and management of evaluation runs.

Bloom’s Four-Stage Pipeline

• Understanding Stage: This stage analyzes the target behavior and provided example conversations to understand the underlying mechanisms and motivation. It summarizes and attributes examples to guide the evaluation suite.
• Ideation Stage: This stage generates diverse evaluation scenarios and variations. Base scenarios are created to elicit the target behavior, followed by variations of each base scenario. The diversity parameter controls the balance between the number of base scenarios and the number of variations per scenario.
• Rollout Stage: This stage executes the generated evaluation scenarios with the target model. This can involve conversational rollouts or simulated environments with tool calls.
• Judgment Stage: This stage evaluates the rollout transcripts, scores the presence of the target behavior, highlights important quotes, and assesses additional user-configured qualities like realism or evaluator-forcefulness. Meta-judgment analyzes qualities across all scenarios.

Getting Started with Bloom

The document outlines the following steps for getting started with Bloom:

• API Key Configuration: Add API keys for the desired model providers to the .env file.
• Virtual Environment Setup: Create and activate a virtual environment using uv and install the required dependencies from requirements.txt.
• Behavior Definition: Define the target behavior in behaviors/behaviors.json with a unique key and description.
• Example Transcripts (Optional): Add example transcripts of the behavior to behaviors/examples/your-behavior-name/, formatted according to the schemas in schemas/.
• seed.yaml Configuration: Customize key parameters in seed.yaml, including the behavior key, example filenames, total evaluation scenarios, and the target model.
• Pipeline Execution: Run the pipeline locally using python bloom.py --debug.
• Transcript Viewing: View the generated transcripts using the interactive viewer: npx @isha-gpt/bloom-viewer --host 0.0.0.0 --port 8080 --dir ./results.

Configuration Parameters

The seed.yaml file contains detailed configuration parameters, categorized as follows:

• Behavior Configuration: Defines the target behavior and provides example conversations.
• Global Parameters: Sets global parameters like temperature, reasoning effort, concurrency, and debug mode.
• Understanding Stage Configuration: Configures the model and token limits for the understanding stage.
• Ideation Stage Configuration: Configures the model, total evaluations, diversity, and token limits for the ideation stage.
• Rollout Stage Configuration: Configures the model, modality, maximum turns, and token limits for the rollout stage.
• Judgment Stage Configuration: Configures the model, token limits, number of samples, and additional qualities for the judgment stage.

Bloom provides a powerful and flexible platform for evaluating LLM behaviors. Its automated suite generation, modular pipeline, and integration with various model providers make it a valuable tool for researchers and developers seeking to understand and mitigate potential biases and unwanted behaviors in LLMs. The detailed configuration options and interactive transcript viewer further enhance the usability and effectiveness of the tool.

Source: bloom