The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts utilize the subprocess module to execute established command-line tools for model evaluation and job management.
Evidence: scripts/inspect_eval_uv.py and scripts/inspect_vllm_uv.py execute the inspect command from the inspect-ai framework.
Evidence: scripts/lighteval_vllm_uv.py executes the lighteval command.
Evidence: scripts/run_eval_job.py and scripts/run_vllm_eval_job.py utilize the hf CLI for job submission and management on Hugging Face infrastructure.
[EXTERNAL_DOWNLOADS]: The skill interacts with well-known technology services to fetch and update model metadata and benchmarks.
Evidence: Uses the huggingface-hub library to load README content and push evaluation results to model cards on Hugging Face.
Evidence: Fetches benchmark data from the Artificial Analysis API (artificialanalysis.ai) using the requests library.
[REMOTE_CODE_EXECUTION]: The evaluation scripts include support for the trust_remote_code parameter, a standard feature in the Hugging Face ecosystem that allows executing custom model code provided in a repository.
Evidence: scripts/inspect_vllm_uv.py, scripts/lighteval_vllm_uv.py, and scripts/run_vllm_eval_job.py allow users to enable this flag for models that require it.
[PROMPT_INJECTION]: The skill processes untrusted content from model READMEs which represents a surface for indirect prompt injection, though it uses structured parsing to mitigate risk.
Ingestion points: scripts/evaluation_manager.py reads README content from various Hugging Face repositories.
Boundary markers: The skill does not use explicit boundary markers when processing README text.
Capability inventory: The skill can update model cards (push_to_hub) and create pull requests on the Hugging Face Hub.
Sanitization: It uses the markdown-it-py library for structured table parsing and validates extracted numeric scores before processing.

hugging-face-evaluation