evaluating-llms-harness

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill documents the use of the lm_eval CLI for running standardized benchmarks like MMLU and GSM8K across different model backends.
  • [DYNAMIC_EXECUTION]: Contains Python examples demonstrating the automation of evaluation workflows using os.system calls within training loops and checkpoint monitoring.
  • [DYNAMIC_EXECUTION]: Explains how to use custom task configurations that leverage the !function YAML tag to dynamically load Python logic from local files for dataset preprocessing and scoring.
  • [DYNAMIC_EXECUTION]: Explicitly references the --allow_code_execution parameter, a standard feature of the evaluation harness used to run model-generated code within a controlled environment for benchmarks such as HumanEval.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 12:48 PM
Security Audit — agent-trust-hub — evaluating-llms-harness