The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches model weights and datasets from established and trusted repositories.
Evidence: Downloads pre-trained models from the Unsloth organization on HuggingFace (unsloth/...) and retrieves safety complaint data from the official NHTSA domain (static.nhtsa.gov).
[COMMAND_EXECUTION]: The documentation provides instructions for using standard machine learning tools.
Evidence: Includes commands for model serving via llama.cpp and experiment tracking using mlflow.
[DATA_INGESTION_RISK]: As a fine-tuning utility, the skill is designed to process external datasets.
Ingestion points: Data is ingested through the datasets library (scripts/demo.py) and direct HTTP requests (scripts/demo_nhtsa.py).
Boundary markers: The skill utilizes standard chat templates (e.g., ChatML, Llama-3) to separate user instructions from input data during training and inference.
Capability inventory: Training operations are restricted to local GPU compute, with output limited to local storage (/tmp/) and local MLflow logging.
Sanitization: The classification logic includes processing steps to clean model outputs (e.g., _strip_thinking) and validate structured JSON responses.

llm-finetuning