The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts inspect_vllm_uv.py and lighteval_vllm_uv.py support a --trust-remote-code parameter. When enabled, this flag allows the underlying libraries (transformers, vllm) to execute custom Python code (modeling and configuration files) directly from the Hugging Face Hub model repository. This is an inherent risk of the Hugging Face ecosystem that the skill facilitates.
[COMMAND_EXECUTION]: All three Python scripts in the scripts/ directory use subprocess.run to execute external CLI tools such as inspect and lighteval. While the scripts use the list-based invocation method which helps prevent shell injection, they pass user-supplied arguments (like model IDs and task names) directly to these subprocesses.
[CREDENTIALS_UNSAFE]: The skill requires and manages the HF_TOKEN environment variable to access gated or private models. The scripts explicitly propagate this token by setting HUGGING_FACE_HUB_TOKEN and HF_HUB_TOKEN in the environment before executing evaluation commands. While this follows standard secret management practices for Hugging Face, it ensures that any code executed (including potentially untrusted remote code) has access to the user's Hugging Face credentials.

huggingface-community-evals