The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the download of the verl framework from the official Volcano Engine GitHub repository and pulls containerized environments from the project's Docker organization. These actions target the official infrastructure of the library.
[COMMAND_EXECUTION]: Provides command-line examples for environment setup, dependency management, and launching distributed training sessions via the library's trainer module.
[INDIRECT_PROMPT_INJECTION]: As a training-focused skill, it describes workflows that ingest large datasets and process model-generated text for reward calculation. It provides template logic for extracting answers from model responses, which is an inherent part of the reinforcement learning from human feedback (RLHF) process.
[DYNAMIC_EXECUTION]: Includes configuration options to load custom reward functions from user-provided local Python scripts, enabling researchers to implement specific scoring logic at runtime.

verl-rl-training