The Agent Skills Directory

[SAFE]: The skill provides legitimate instructions and code samples for LLM evaluation. The Python code snippets demonstrate the use of well-known libraries such as NLTK, Transformers, and Scikit-learn for calculating standard metrics like BLEU, ROUGE, and BERTScore.- [EXTERNAL_DOWNLOADS]: Fetches pre-trained models from a well-known service's repository for BERTScore and NLI-based groundedness metrics. These downloads are part of the standard functionality for the libraries used (SKILL.md).- [INDIRECT_PROMPT_INJECTION]: The skill's implementation patterns for LLM-as-judge and automated metrics involve ingesting untrusted data (model outputs and test inputs) into the evaluation pipeline. This represents an attack surface where malicious data could attempt to influence the evaluator.
- Ingestion points: Untrusted data enters the context via the test_cases list and response variables in the llm_judge_quality and compare_responses functions (SKILL.md).
- Boundary markers: The provided prompt templates use structural delimiters (e.g., 'Question:', 'Response:') to separate data from instructions.
- Capability inventory: The skill uses standard data processing and calls to the OpenAI API; it does not perform file-writes or dangerous system calls.
- Sanitization: No explicit sanitization or escaping of the processed model outputs is shown in the educational snippets.

llm-evaluation