The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill involves loading external datasets from well-known sources like Hugging Face Hub for model training. While these datasets contain content generated by third parties, the skill handles them through standard machine learning pipelines where the data is used for optimization rather than direct execution. The risk is inherent to the purpose of model alignment.\n
Ingestion points: The skill uses datasets.load_dataset in SKILL.md and templates/basic_grpo_training.py to fetch training data.\n
Capabilities: The skill includes model training routines and filesystem operations for saving results, which are restricted to the local environment.\n- [COMMAND_EXECUTION]: The instructions guide the user to install necessary Python packages and run training scripts via the command line. These commands are essential for setting up the environment and executing the fine-tuning workflows described in the skill.\n
Commands: Includes pip install for dependencies and CLI usage for trl and python scripts.\n- [EXTERNAL_DOWNLOADS]: The skill fetches pre-trained models and datasets from Hugging Face, an industry-standard repository. These resources are from established organizations and are standard components of the machine learning ecosystem.\n
Sources: References models and datasets from the trl-lib and Qwen repositories on Hugging Face.

fine-tuning-with-trl