The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads model weights from Hugging Face and a vehicle complaint dataset from the official NHTSA website. These are reputable and well-known sources for machine learning research and public safety data.
[INDIRECT_PROMPT_INJECTION]: The skill processes external text datasets which could theoretically contain malicious instructions. However, it implements security best practices by using the train_on_responses_only feature to mask prompts during training, ensuring the model only learns from designated response fields. It also includes functions to clean model output and uses structured JSON parsing for multi-target classification tasks.
Ingestion points: scripts/demo.py (via load_dataset) and scripts/demo_nhtsa.py (via requests.get).
Boundary markers: Instruction and response masking tokens are explicitly defined and applied using train_on_responses_only in both demo scripts.
Capability inventory: The skill performs network operations to fetch data/models and writes training artifacts to local temporary directories.
Sanitization: The _strip_thinking function in both demo scripts removes internal reasoning blocks, and _parse_prediction in demo_nhtsa.py uses JSON schema validation.

llm-finetuning