The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads pre-trained models and datasets from Hugging Face Hub, which is a well-known and trusted service. Examples include the Qwen model family and the Capybara dataset.
[COMMAND_EXECUTION]: Provides instructions for running the trl and accelerate command-line tools to execute model training and serving workflows.
[DYNAMIC_EXECUTION]: The skill utilizes user-defined Python functions, such as reward functions in the GRPOTrainer, which are executed locally during the reinforcement learning loop.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data from Hugging Face repositories for model training, creating a potential surface for instructions embedded in training data.
Ingestion points: External datasets are loaded via load_dataset calls in SKILL.md.
Boundary markers: The instructions do not specify delimiters or constraints for the content of the training datasets.
Capability inventory: The workflow involves shell command execution via the trl CLI and model training subprocesses.
Sanitization: No explicit validation or filtering of the dataset content is described.

trl-training