AgileRL Arena v1.0 Released

AgileRL has released AgileRL Arena v1.0 as an independent PyPI package. This SDK/CLI has no torch dependencies and allows defining reinforcement learning (RL) runs with the same configuration on local or managed clusters via YAML manifest files.

Key Features

• Manifest-based execution: Define algorithms, environments, networks, mutation, and selection strategies in YAML. Validation using Pydantic returns errors immediately for invalid configurations before training starts.
• LLM fine-tuning support: Accepts Hugging Face model IDs and LoRA configurations to perform fine-tuning. Supports rollout environments that apply reward functions to model outputs, as well as supervised learning datasets using objective: sft or preference.
• Easy local execution: Start training on a local GPU with two lines of code using the LocalTrainer class, without an account.
• Evolutionary hyperparameter optimization (HPO): Supports training a population instead of a single agent, periodically mutating architectures and hyperparameters to select optimal individuals.

Changes in v1.0

• Algorithm specifications have changed to a strategy selection method based on paradigm flags, simplifying code branching.
• Configuration files can be validated before submission using the arena manifest validate and arena manifest schema commands.
• The PEFT library has been updated to version 0.20, automatically excluding LoRA application to out_proj and conv1d layers in Mamba models.