[slikhite-1]

This PR implements an end-to-end example for training a model on multi-turn tool calling using GRPO with the BFCLv3 dataset. The example leverages preprocessed data uploaded to HuggingFace, where preprocessing includes adding tool descriptions and formatting. It also provides evaluation scripts to compare trained and untrained models, along with a README that documents the example in detail, and results on Qwen 2.5 3B

Summary by CodeRabbit

• New Features
  • Introduced multi-turn tool-calling environment with reward-based rollouts.
  • Added support for BFCL v3 multi-turn datasets for training and evaluation.
  • Provided new built-in tools (filesystem, math, messaging, ticketing, trading, social, vehicle control).
  • Enabled multi-turn evaluation and GRPO training via new scripts and configs.
  • Improved logging with serialized environment metadata.
• Documentation
  • Added a comprehensive guide for multi-turn tool calling; linked from the guides index.
• Chores
  • Included a new dependency for JSON Lines data processing.