The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides installation commands that pipe remote scripts directly into the shell, such as curl -fsSL https://unsloth.ai/install.sh | sh and irm https://unsloth.ai/install.ps1 | iex. These commands originate from the official domain of the library being documented.\n- [EXTERNAL_DOWNLOADS]: The skill references and provides commands to download code and configuration from official repositories, including github.com/unslothai/unsloth and github.com/ggerganov/llama.cpp. These are standard sources for the technology described.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to help users fine-tune models using external datasets, which creates a potential surface for indirect prompt injection.\n
Ingestion points: Data is loaded from HuggingFace Hub or local files (JSON, CSV, PDF) as described in references/dataset-formatting.md.\n
Boundary markers: The provided training templates do not include explicit instructions or markers to distinguish between training instructions and potentially malicious content within the dataset entries.\n
Capability inventory: The training environment outlined in references/training-setup.md has the capability to write files to the local system (model outputs) and perform network operations (pushing to the HuggingFace Hub).\n
Sanitization: The skill does not specify any sanitization or validation routines for the data processed during the fine-tuning workflow.

unsloth-expert