hud-vf-gym 0.1.1

HUD environments for Reinforcement Learning through Verifiers

hud-vf-gym

Verifiers Adapter for HUD environments - bridges Verifiers RL framework with HUD's MCP infrastructure for training and evaluating agents.

Training Support

Verifiers' GRPOTrainer does not support multimodal training as of now, you can use an experimental trainer for single-turn environments (with single prompt image due to transformer's limitations). Multi-turn multimodal support is WIP.

Prerequisites

• Python >=3.12
• HUD API key from https://app.hud.so
• Environment variables:

  export HUD_API_KEY="your-api-key"
  export OPENAI_API_KEY="your-key"  # or ANTHROPIC_API_KEY
  

Quick Start

import verifiers as vf

# Load environment with HUD taskset and config
env = vf.load_environment(
    env_id="hud-vf-gym",
    taskset="hud-evals/2048-taskset",  # HuggingFace dataset
    config_path="./configs/2048.yaml",  # Environment config
    num_tasks=10
)

Documentation

For comprehensive usage, examples, and configuration details, see: hud-python/rl README

The main documentation covers:

• Running evaluations with various models
• Training agents with GRPO
• Creating custom environments and configs
• Dataset format and creation
• Action mappings and tool configuration
• Troubleshooting guide

License

MIT