The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the litgpt package from PyPI and fetches pretrained model weights from well-known repositories such as HuggingFace, Microsoft, and Google. These are standard operations for machine learning development targeting official and trusted sources.
[COMMAND_EXECUTION]: The skill provides instructions to use the litgpt command-line interface and local Python scripts for training, fine-tuning, and model conversion. These operations are essential to the skill's primary purpose and involve standard data processing and model management workflows.
[INDIRECT_PROMPT_INJECTION]: The skill supports processing user-provided datasets for fine-tuning operations.
Ingestion points: Custom datasets are loaded from local files such as data/my_dataset.json (documented in SKILL.md).
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are mentioned in the provided examples.
Capability inventory: The skill possesses capabilities to write files (checkpoints, logs), execute shell commands via the CLI, and perform network requests to download models (documented in SKILL.md and references/*.md).
Sanitization: No explicit sanitization or validation of the training data content is described, representing a surface where malicious instructions in a dataset could influence model training or the agent's behavior during the fine-tuning process.

model-architecture-litgpt