The Agent Skills Directory

[DYNAMIC_EXECUTION]: The function run_test_cases in examples/reward_functions_library.py uses the exec() function to execute model-generated code blocks during the reward calculation phase. While intended for evaluating code generation tasks, the use of exec() on untrusted model outputs without isolation (such as containers or sandboxes) poses a risk of arbitrary code execution on the training host. The skill includes a documentation warning suggesting the use of sandboxing for production environments.- [INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by ingesting external datasets and model completions.
Ingestion points: Data is ingested from external sources via load_dataset in templates/basic_grpo_training.py and from model-generated completions in examples/reward_functions_library.py.
Boundary markers: The implementation lacks explicit boundary markers or instructions to the model to ignore potential injections within the training data or generated responses.
Capability inventory: The training environment includes capabilities to execute code (exec() in examples/reward_functions_library.py) and write to the local file system (trainer.save_model in templates/basic_grpo_training.py).
Sanitization: No input validation, sanitization, or sandboxing is implemented in the provided reward function examples to protect against malicious payloads in completions or datasets.- [EXTERNAL_DOWNLOADS]: The skill references and downloads resources from well-known services and organizations, including Hugging Face for model weights (transformers) and datasets (datasets), and Weights & Biases (wandb) for experiment tracking. These operations are standard for the described machine learning workflow.

grpo-rl-training