The Agent Skills Directory

[COMMAND_EXECUTION]: The script examples in references/deployment-optimization.md utilize the subprocess module to call external model conversion and quantization tools associated with the llama.cpp project. This is standard industry practice for preparing models for efficient inference and does not involve arbitrary or unsafe command injection.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external training datasets in JSONL and Parquet formats through libraries like datasets. To mitigate risks from malicious or low-quality data content, the skill provides extensive validation logic, including token length verification, regex-based quality filtering (e.g., removing AI refusal patterns), and deduplication routines (Exact and Fuzzy MinHash LSH).
[EXTERNAL_DOWNLOADS]: The Python examples include functionality to download pre-trained model weights, tokenizers, and benchmarking datasets from official sources such as the Hugging Face Hub (e.g., Meta Llama models, WikiText dataset). It also supports logging metrics to established services like Weights & Biases and pulling official Docker images for Text Generation Inference (TGI).

fine-tuning-expert