The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references downloading models and datasets from Hugging Face and using specialized services like Weights & Biases (WandB), MLflow, and Trackio for observability. These are well-known services or trusted organizations in the machine learning ecosystem. The usage is gated by user-provided environment variables (e.g., WANDB_API_KEY, HF_TOKEN).
[COMMAND_EXECUTION]: The skill provides examples of shell commands for platform configuration (e.g., evo config set) and model serving (e.g., vllm serve). These are standard operational commands within the intended context of the skill and do not involve unauthorized privilege escalation or persistence mechanisms.
[DATA_EXFILTRATION]: The skill facilitates the transmission of training metrics (loss, learning rate, etc.) to external tracking services. This is a standard practice for experiment monitoring, is explicitly disclosed to the user, and requires valid API keys provided via environment variables.
[PROMPT_INJECTION]: The skill includes instructions for managing the agent's decision-making process (e.g., "Override anything else against the gate"). These are technical guardrails for model training quality and do not represent attempts to bypass safety filters or ignore system instructions for malicious purposes.

finetuning