The Agent Skills Directory

[COMMAND_EXECUTION]: The convert_to_gguf.py script uses subprocess.run to execute commands for cloning repositories, installing packages with pip, and configuring builds with cmake. The script implements defensive regex validation via require_hf_repo_id and safe_filename_component to verify that Hugging Face repository IDs and local path components are properly formatted before use, which helps mitigate command injection risks.
[DYNAMIC_EXECUTION]: The model conversion process involves the runtime compilation of the llama.cpp C++ library using cmake. The skill builds the llama-quantize binary at runtime and subsequently executes it to generate model files in different quantization formats.
[EXTERNAL_DOWNLOADS]: The skill fetches the llama.cpp source code from GitHub and references a dataset inspection script hosted on Hugging Face. These downloads target well-known and reputable services standard to the machine learning ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets from the Hugging Face Hub, creating a potential surface for indirect prompt injection.
Ingestion points: Data is ingested via load_dataset in training templates such as scripts/train_sft_example.py and scripts/unsloth_sft_example.py.
Boundary markers: The scripts do not explicitly use boundary markers or specific instructions to ignore embedded prompts within the dataset text.
Capability inventory: The skill can execute cloud-based jobs via the hf_jobs tool and perform local file writes and binary execution during the GGUF conversion process.
Sanitization: While repository names are validated, the raw content of the ingested datasets is not sanitized for malicious natural language instructions.
[PRIVILEGE_ESCALATION]: The skill documentation and scripts include commands to install system-level packages using apt-get. While these operations typically require elevated privileges, they are intended for the configuration of the Hugging Face Jobs ephemeral Docker environment rather than the local host.

hugging-face-model-trainer