The Agent Skills Directory

[SAFE]: The skill exclusively provides instructional content and code examples for using the well-known bitsandbytes and HuggingFace transformers libraries. All operations (quantization, training, inference) are standard machine learning workflows with no evidence of malicious intent or hidden capabilities.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing official, well-known Python packages from standard registries (PyPI) and references official documentation from HuggingFace and the bitsandbytes foundation. These resources are from established entities in the AI research community.
[COMMAND_EXECUTION]: Shell commands are restricted to standard package installation (pip install) and CUDA version checking (nvcc --version). No arbitrary or dangerous command execution was detected.
[DATA_EXPOSURE]: The skill uses local paths for output and offloading (e.g., ./qlora-output, ./offload) and does not access sensitive system files or environment credentials. It demonstrates standard practices for model and adapter storage.
[INDIRECT_PROMPT_INJECTION]: While the skill involves processing instruction datasets for fine-tuning, the documentation focuses on the infrastructure and quantization configuration. It includes standard warnings regarding accuracy trade-offs and provides clear technical checklists to ensure user oversight of the process.

quantizing-models-bitsandbytes