The Agent Skills Directory

[SAFE]: The skill provides documentation and code examples for memory-efficient fine-tuning using QLoRA. All code snippets use standard libraries such as transformers, peft, and bitsandbytes. All external references target well-known services or trusted organizations.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a standard training workflow that involves ingesting external data, which is an inherent attack surface for indirect prompt injection. * Ingestion points: SKILL.md includes code to load the tatsu-lab/alpaca dataset using the datasets library. * Boundary markers: The training prompt template uses standard text delimiters (e.g., ### Instruction:, ### Response:) which provide organizational structure but are not robust security boundaries. * Capability inventory: The skill utilizes model training capabilities (trainer.train()), file system writes for saving models and adapters (model.save_pretrained()), and environment variable configuration (export LD_LIBRARY_PATH). * Sanitization: No explicit data sanitization or filtering is performed on the training dataset within the provided examples, which is consistent with the primary purpose of a technical demonstration.

qlora