The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for training large language models using external datasets provided in Parquet format. This presents a potential surface for indirect prompt injection if the dataset contains adversarial content designed to influence the training process or the behavior of the resulting model.
Ingestion points: Training data files specified in the configuration (e.g., data.train_files in SKILL.md).
Boundary markers: The provided configuration snippets do not explicitly define boundary markers or instruct the agent to ignore instructions embedded within the data.
Capability inventory: The skill utilizes command-line execution for the trainer (python3 -m verl.trainer.main_ppo) and allows for file system access to training data.
Sanitization: No explicit sanitization or validation of the dataset content is described in the provided workflows.
[DYNAMIC_EXECUTION]: The training framework supports the execution of custom Python logic for reward calculation by loading scripts from user-defined local paths. While this is a standard requirement for reinforcement learning tasks, it involves runtime execution of code that is not part of the core library.
Evidence: The custom_reward_function.path parameter in references/api-reference.md allows users to specify an arbitrary local Python file to be executed during the reward phase.

verl-rl-training