Normalize action string for comparison.

Normalize answer string for comparison.

Merge new generated tasks with the original training dataset.

Release trainer resources.

Compare extracted actions with expected actions.

Evaluate agent performance on the current training dataset. Returns: Dictionary mapping prompt -> {'scores': [...], 'gt': ground

Extract action calls from text.

Extract the last LaTeX \\boxed{...} content from text.

Extract tool calls from <tool_call> tags.

Filter out prompts that are too easy (always solved) or too hard (never solved). The GenEnv algorithm focuses on prompts at the boundary

Generate new training tasks using the Environment LLM. >>> USER CUSTOMIZATION REQUIRED <<< Modify the prompt template and parsing lo

Extract the gold answer from ground truth.

Check if predicted call matches any gold call.

Parse ground truth into list of actions.

Compute reward for a single generated response. >>> USER CUSTOMIZATION REQUIRED <<< Replace this method with your do

Main entry point for GenEnv training.

Main entry point for GenEnv co-training. This function orchestrates the alternating training between Agent and Environment. Arg

Compute rewards for a batch of generated responses. Args: data: DataProto containing prompts, responses, and gro

Initialize the reward manager. Args: tokenizer: HuggingFace tokenizer for decoding sequences num_exa

Initialize the Environment Generator. Args: model_path: Path to the Environment LLM (e.g., Qwen2.5-7B-Instruct)

Check if predicted tool calls match ground truth. Expected ground_truth format: {"name": "api_name", "parameters": {

Check if predicted actions match ground truth. Expected actions format: - type(bid=123, value=text, press_enter=True

Generate new tasks from the given prompts. Args: prompts: List of formatted prompts for task generation

Ray remote task for GenEnv co-training.

Ray remote task for standard GRPO training (without co-training).