The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends installing several standard Python packages for benchmarking, including lm-eval, vllm, and human-eval.
[COMMAND_EXECUTION]: The instructions provide multiple examples of running the lm_eval CLI tool. Additionally, documentation in SKILL.md includes Python snippets that utilize os.system for automating evaluation workflows and running shell scripts.
[DYNAMIC_EXECUTION]: The skill demonstrates how to enable code execution for specific tasks, such as HumanEval, via the --allow_code_execution flag. Furthermore, references/custom-tasks.md provides examples of custom task utilities that use the subprocess module to execute code generated or referenced by the model.
[INDIRECT_PROMPT_INJECTION]: The skill documents an evaluation process that ingests and processes external data (benchmarks and datasets), creating an indirect prompt injection surface.
Ingestion points: External data enters the context when the tool reads benchmark datasets from HuggingFace or local files (JSONL, CSV) as described in SKILL.md and references/custom-tasks.md.
Boundary markers: The workflow uses standardized Jinja2 prompt templates to structure inputs, but lacks explicit sanitization instructions for adversarial data within those benchmarks.
Capability inventory: The skill provides instructions for CLI execution, shell command interpolation in Python scripts, and explicit code execution via the benchmarking tool's internal flags.
Sanitization: There is no mention of escaping or filtering content from external benchmarks; the process relies on the isolated execution environment of the lm-evaluation-harness tool.

evaluating-llms-harness