The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documents procedures for downloading large language model weights and datasets from HuggingFace, a well-known service in the machine learning community. These downloads target official organizations such as Meta (Llama series) and reputable researchers (e.g., Tim Dettmers).
[COMMAND_EXECUTION]: Instructions are provided for setting up the development environment using standard pip commands to install well-known libraries from the Python Package Index (PyPI), including bitsandbytes, transformers, and accelerate.
[PROMPT_INJECTION]: The skill contains workflows that ingest external datasets, which constitutes an indirect prompt injection surface. However, since the skill is designed for developer instruction rather than autonomous agent execution, the risk is negligible.
Ingestion points: Training data is loaded from the HuggingFace Hub in the references/qlora-training.md file.
Boundary markers: Code snippets utilize standard instruction-tuning delimiters (e.g., ### Human:, ### Assistant:) which help distinguish between data and control instructions.
Capability inventory: Capabilities documented are limited to model fine-tuning, memory profiling, and saving model adapters to the local filesystem. No arbitrary shell execution or network exfiltration capabilities are present.
Sanitization: No explicit sanitization or filtering of dataset content is demonstrated, which is typical for tutorials focused on the training process itself.

quantizing-models-bitsandbytes