The Agent Skills Directory

[SAFE]: The skill serves as a documentation and command-reference tool for the Hugging Face TRL (Transformers Reinforcement Learning) library, facilitating model fine-tuning tasks such as SFT, DPO, and GRPO.
[EXTERNAL_DOWNLOADS]: The instructions reference models and datasets hosted on the Hugging Face Hub (e.g., trl-lib/Capybara, Qwen/Qwen2-0.5B). These resources originate from well-known organizations and are accessed through standard library calls, posing no inherent security risk to the agent environment.
[COMMAND_EXECUTION]: The skill provides standard shell command templates for the trl and accelerate CLI tools. These commands are localized to machine learning workflows and include guidance on memory management and distributed training configurations.
[INDIRECT_PROMPT_INJECTION]: While the skill involves processing external datasets, it includes explicit security boundaries in its 'Anti-Patterns' section, instructing the agent to never treat external content, logs, or generated output as trusted instructions.
[CREDENTIALS_UNSAFE]: The skill mentions authentication via hf auth login for accessing gated models. This is a standard user-initiated security procedure for the Hugging Face ecosystem and does not involve hardcoded secrets or automated exfiltration patterns.

trl-training