unsloth-finetuning
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads large language models and training datasets from the Hugging Face Hub (e.g., unsloth/Qwen3-8B and mlabonne/FineTome-100k).
- [REMOTE_CODE_EXECUTION]: The model.save_pretrained_gguf method downloads and compiles the llama.cpp toolchain from a remote source on its first invocation.
- [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: External training data is ingested via the load_dataset function in SKILL.md.
- Boundary markers: The skill does not implement automated boundary markers or prompt delimiters to isolate instructions embedded within the training data.
- Capability inventory: The skill possesses file-writing capabilities, network access, and the ability to compile and execute external code.
- Sanitization: No technical sanitization or filtering of dataset content is performed beyond recommending manual review.
- [DYNAMIC_EXECUTION]: Exporting models to GGUF format triggers the runtime compilation of external C++ source code (llama.cpp).
Audit Metadata