The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/convert_to_gguf.py utilizes subprocess.run to call system utilities including git for cloning repositories, cmake for build configuration, and pip for installing dependencies derived from external sources.
[REMOTE_CODE_EXECUTION]: The model conversion process involves cloning the llama.cpp repository from GitHub and executing its internal Python scripts and compiled binaries. Additionally, the skill uses the trust_remote_code=True parameter when loading models, allowing for the execution of custom code provided within the model's repository.
[DATA_EXFILTRATION]: The skill is designed to upload trained model weights, configuration files, and training logs to the Hugging Face Hub. This process requires the use of a user-supplied HF_TOKEN with write permissions passed through the environment.
[PROMPT_INJECTION]: The skill ingests and processes external datasets from the Hugging Face Hub, which could contain adversarial instructions intended to influence the training process or the agent's behavior.
Ingestion points: Data is loaded via datasets.load_dataset() in several training templates, such as scripts/train_sft_example.py and scripts/train_dpo_example.py.
Boundary markers: No specific delimiters or validation logic is implemented to identify or ignore instructions embedded within the dataset text fields.
Capability inventory: The skill can execute shell commands locally via subprocess and submit scripts for remote execution on managed infrastructure using the hf_jobs tool.
Sanitization: The skill performs standard data loading and formatting without specific sanitization or filtering of the content retrieved from external datasets.

hugging-face-model-trainer