The Agent Skills Directory

[COMMAND_EXECUTION]: The documentation in references/deployment-optimization.md includes Python functions that utilize the subprocess module to execute external binaries and scripts. Specifically, the export_to_gguf function calls the llama-quantize tool and a conversion script from a path determined by the LLAMA_CPP_PATH environment variable. While typical for model quantization workflows, this involves system-level command execution.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and processing external datasets (JSONL, Parquet) for model training, which presents a surface for indirect prompt injection. Mandatory Evidence Chain:
Ingestion points: Dataset loading logic in SKILL.md and references/dataset-preparation.md.
Boundary markers: The skill does not provide specific instructions or delimiters to isolate embedded commands in training data.
Capability inventory: The skill has the ability to execute shell commands via subprocess.run and perform network operations via transformers and openai libraries.
Sanitization: Basic quality filters in references/dataset-preparation.md search for and remove common AI refusal patterns.
[EXTERNAL_DOWNLOADS]: The skill uses standard libraries to download pre-trained model weights, tokenizers, and datasets from the Hugging Face Hub. It also supports evaluation using the OpenAI API. These operations are essential for the skill's stated purpose of model fine-tuning.

fine-tuning-expert