atropos/environments/community/README.md

Community Environments

This directory is home to community-contributed training environments for Atropos. Environments submitted by the community will be placed here after an initial code review.

Note: Environments in this directory are pending full testing and integration. While they have passed a basic code check, they may not yet have been rigorously validated on our compute cluster.

Contributing Your Environment

We encourage you to contribute your own RL environments! When developing a new environment, please follow these guidelines:

1. Create your environment in this environments/community/ subdirectory. This helps us keep new submissions organized.
2. Preferred Import Style: We prefer that you treat your environment's directory as the package root for imports within your environment code. For example, if you need to import SomeClass, you can do so directly:

   from some_file_in_my_env import SomeClass
   

   This helps maintain consistency and makes it easier to integrate your environment.

Environment Standards

Community environments should:

• Include clear documentation and setup instructions
• Specify all dependencies in requirements files
• Provide example configurations and usage
• Follow the AtroposBaseEnv pattern for consistency
• Include appropriate error handling and validation

Submission Process

To contribute a new environment to the community collection:

1. Fork the repository and create a new branch
2. Add your environment to this community/ directory
3. Include comprehensive documentation:
   • README with setup instructions
   • Requirements file for dependencies
   • Example usage and configuration
4. Follow naming conventions:
   • Use descriptive directory names for complex environments
   • Single file environments should have descriptive names
5. Test thoroughly before submitting
6. Submit a pull request with a clear description

Once your environment is ready, please follow the guidelines in our main CONTRIBUTING.md to submit your contribution.

---

Available Environments

1. Lean Proof Environment (lean_proof_env/)

Author: GabinFay Purpose: Testing Language Learning Models (LLMs) on Lean theorem proving tasks

A comprehensive environment for evaluating LLMs on formal mathematical reasoning using the Lean theorem prover. Features include:

• Support for custom problem datasets or MiniF2F benchmark
• Integration with Lean 4 theorem prover
• Configurable difficulty levels and problem sets
• Automated proof validation

Requirements: Lean 4 installation, OpenAI API key

2. Router Environment (router_env/)

Author: GabinFay Purpose: Multi-agent routing and coordination system

A sophisticated environment for testing agent routing and coordination capabilities. Includes:

• Multiple specialized agents (calendar, contact, Gmail, telephony, etc.)
• Model Contextualized Protocol (MCP) tools integration
• Spotify, Google Maps, and Perplexity integrations
• Complex multi-turn conversation handling

Features:

• Telephony agent with inbound/outbound call handling
• Calendar and contact management
• Memory and calculation agents
• Router agent for intelligent task delegation

3. Philosophical RLAIF Environment (philosophical_rlaif_env.py)

Author: GabinFay Purpose: Reinforcement Learning from AI Feedback (RLAIF) for philosophical reasoning

An environment focused on training models for deep philosophical inquiry and reasoning. Features:

• Deep thinking prompts with systematic reasoning processes
• Preference learning for philosophical depth and nuance
• Multi-perspective analysis and assumption questioning
• Evaluation of response quality for philosophical discussions

Capabilities:

• Generates paired responses for preference comparison
• Uses judge models to evaluate philosophical depth
• Tracks preference consistency and reasoning quality
• Supports WandB logging for training insights

4. Playwright Agent Environment (playwright_agent_env.py)

Author: erikqu Purpose: Web automation and browser interaction for LLM agents

A comprehensive environment for training LLMs to interact with web pages through browser automation. Features:

• Playwright-based browser control with headless operation
• Screenshot-based visual input for LLM decision making
• JSON-based action commands (navigate, click, type, finish)
• Video recording of browser sessions for evaluation
• Google Gemini integration for success evaluation

Capabilities:

• Loads tasks from WebVoyager dataset or custom task definitions
• Supports development mode for testing without LLM calls
• Automatic reward computation based on success and efficiency
• Comprehensive error handling and fallback mechanisms
• Integration with Atropos training pipeline

Requirements: Playwright, optional Google Gemini API for evaluation

5. Metric Card Generator Environment (metric_card_generator/)

Author: vivek100 Purpose: Structured JSON generation for AI model evaluation dashboards

A comprehensive environment for training LLMs to generate well-structured JSON configurations for Metric Card UI components. Features:

• Closed-loop generation, evaluation, and visualization pipeline
• Schema validation for JSON metric card configurations
• Multi-dimensional evaluation (validity, compliance, semantic quality)
• Support for various business domains and metric types
• WandB integration for performance tracking

Capabilities:

• Generates metric cards for diverse business contexts (e-commerce, finance, healthcare, etc.)
• Validates JSON structure against predefined schemas
• Evaluates semantic quality and formatting consistency
• Provides training data extraction and filtering utilities
• Includes visualization tools for score distribution analysis

Components:

• metric_card_generator.py: Main environment implementation
• extract_metric_training.py: Training data extraction utility
• trainingDataScript.py: Dataset creation from collected examples
• show_score_distribution.py: Performance analysis visualization

Requirements: Pydantic, tqdm

6. UFC Prediction Environment (ufc_prediction_env/)

Author: edmundman Repository: UFC_FIGHT_PREDICTOR Purpose: UFC fight prediction with entertaining TTS-ready commentary generation

A creative environment that transforms traditional fight prediction into engaging entertainment by generating dynamic, broadcast-style UFC fight commentary. Features both text-based and image-based prediction modes:

Text-Based Predictor (ufc_server.py):

• Uses comprehensive fighter statistics (wins/losses, physical attributes, performance metrics)
• Generates dramatic fight commentary with commentator personalities
• TTS-ready output with natural speech patterns and emphasis markers
• Statistical analysis wrapped in entertaining storytelling

Image-Based Predictor (ufc_image_env.py):

• Multimodal prediction using fighter profile images
• Visual analysis transformed into engaging commentary
• Base64 image encoding for API compatibility
• Creates dramatic narratives from fighter appearances

Key Features:

• Entertainment-first approach with broadcast-style commentary
• Direct TTS integration compatibility (designed for models like DIA)
• Dramatic elements including commentator phrases and pauses
• Proper formatting for voice synthesis applications
• Comprehensive scoring system for prediction accuracy and entertainment value

Data Components:

• fighter_stats.csv: Detailed fighter statistics and performance metrics
• large_dataset.csv: Sample historical fight data (799 records from original 7,440)
• fighter_images/: Profile images for visual-based predictions
• get_images.py: Web scraping utility for fighter image collection

Note: The included dataset is a sample for demonstration. The full dataset (7,440 fight records) is available in the original UFC_FIGHT_PREDICTOR repository.

Additional Tools:

• ufc_predictor_ui.py: Flask-based web interface for interactive predictions
• Video demonstrations and example runs available
• WandB integration for training tracking

Requirements: PIL, OpenAI API, Flask (for UI), BeautifulSoup4 (for image scraping)

---

Support

For questions or issues with community environments:

• Check the individual environment's README first
• Open an issue in the main repository
• Tag the environment author if possible

These environments are community contributions and may have different maintenance levels and support compared to core Atropos environments.