The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the installation of standard machine learning libraries (auto-gptq, transformers, optimum, peft) and fetches models/datasets from HuggingFace and GitHub. All resources originate from well-known services or trusted organizations.
[COMMAND_EXECUTION]: Provides standard pip install commands for environment setup and shell commands for running inference servers via Docker, which are routine for this skill's technical domain.
[DATA_EXPOSURE]: Includes instructions for pushing quantized models to HuggingFace using model.push_to_hub("username/..."). This follows standard developer workflows and does not hardcode or exfiltrate sensitive credentials.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (e.g., C4, ShareGPT) for model calibration. While this is an ingestion point for untrusted data, the skill uses standard ML loaders (datasets.load_dataset) and the risk is limited to model bias/quality rather than system exploitation.

gptq