The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the installation script for the 'uv' package manager from 'astral.sh'. Astral is a well-known service provider in the Python development ecosystem, and 'uv' is widely used for performant package management.
[COMMAND_EXECUTION]: Employs 'pre_cmd' blocks to automate the setup of evaluation environments. This includes installing system dependencies (such as Apptainer for sandboxing) and Python libraries required for specific benchmarks. This is a core functional requirement of the NeMo Evaluator platform to ensure consistent test environments.
[INDIRECT_PROMPT_INJECTION]: As an LLM evaluation tool, the skill naturally ingests and processes untrusted output from models under test. The skill mitigates this risk by using 'adapter_config' to strip reasoning traces before scoring and by executing agentic tasks within isolated Apptainer containers.
[DYNAMIC_EXECUTION]: The skill uses dynamic environment setup during the evaluation lifecycle, such as checking out specific commits of the NeMo Gym repository and installing dependencies on-the-fly. These operations target official project repositories and are used to ensure the evaluation harness matches the expected version for a given benchmark.
[CREDENTIALS_UNSAFE]: The skill manages sensitive API keys (e.g., HF_TOKEN, INFERENCE_API_KEY) via environment variables and '.env' files. It provides clear guidance to users on safe secret management, including explicit warnings against editing sensitive files within the agent's view to prevent leakage into session logs.

evaluation