The Agent Skills Directory

[SAFE]: The skill provides high-quality educational content and functional Python code snippets for fine-tuning Large Language Models. All code uses established, open-source libraries from reputable sources.
[DATA_EXPOSURE]: No hardcoded credentials, API keys, or sensitive file paths (such as SSH keys or cloud configurations) were detected. The skill correctly instructs users to handle model weights and adapters locally or through official platforms.
[EXTERNAL_DOWNLOADS]: The skill references standard datasets and models hosted on Hugging Face (e.g., trl-lib/Capybara, Qwen/Qwen2.5-7B-Instruct). These are well-known repositories within the machine learning community and are used for legitimate training purposes.
[COMMAND_EXECUTION]: The provided code snippets are limited to Python-based training logic using the trl and transformers APIs. There is no evidence of arbitrary shell command execution, privilege escalation, or persistence mechanisms.
[PROMPT_INJECTION]: The instructions are clear, descriptive, and do not contain patterns aimed at bypassing AI safety guardrails or overriding system prompts.
[INDIRECT_PROMPT_INJECTION]: The skill inherently processes external data (training datasets) which is an entry point for untrusted content. However, the skill provides mitigation guidance, such as using the training-data skill for corpus cleaning and validation, and applying chat templates to maintain proper sequence boundaries.

finetuning