unsloth
Pass
Audited by Gen Agent Trust Hub on Apr 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install third-party libraries (unsloth, vllm, transformers) and clone repositories (unslothai/unsloth, ggerganov/llama.cpp). These sources are well-established and reputable projects within the machine learning community used for model quantization and optimization.\n- [COMMAND_EXECUTION]: The documentation includes standard shell commands for package management (pip, uv) and system-level build tools (apt-get, cmake) used in model deployment workflows. These commands are presented as manual steps for the user to configure their environment.\n- [PROMPT_INJECTION]: The skill inherently processes external datasets as part of the model training and fine-tuning process. This is an expected and primary function of the skill.\n
- Ingestion points: datasets are loaded using the Hugging Face
load_datasetlibrary in SKILL.md and references/embeddings.md.\n - Boundary markers: Chat templates and tokenizers are used to manage instruction boundaries.\n
- Capability inventory: The training process involves significant compute and file-system writes to save checkpoints.\n
- Sanitization: Input data is processed via standard tokenizers and model-specific chat templates.
Audit Metadata