The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the lm-eval harness via subprocess.Popen in scripts/eval_dataset.py. This is a standard and isolated method for running the benchmarking tool.
[EXTERNAL_DOWNLOADS]: The skill fetches benchmark datasets (e.g., GSM8K, MMLU) from Hugging Face and installs required Python packages using the Pixi environment manager. These downloads target well-known and reputable services essential for the skill's primary function.
[DATA_EXFILTRATION]: The skill reads the OPENAI_API_KEY from the environment and includes it in the Authorization header for requests sent to the user-specified --base-url. This is the standard mechanism for authenticating with OpenAI-compatible model endpoints.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from external datasets for the purpose of model evaluation. While this creates a data ingestion surface, it is a necessary component of model benchmarking.
Ingestion points: Benchmark samples are loaded from Hugging Face in scripts/_eval_tasks.py.
Boundary markers: Not explicitly defined for dataset samples, as the tool relies on automated scoring against ground-truth labels.
Capability inventory: The skill has access to network operations (for API calls) and subprocess execution (for the evaluation harness).
Sanitization: The skill processes dataset content as provided by Hugging Face to ensure the integrity of standard benchmark results.

eval-model