The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.check_output and subprocess.Popen in several scripts (sft_smoke.py, verify_export.py, test_unsloth_preflight.py) to interface with nvidia-smi and execution of target runtimes like llama.cpp. These are implemented with strict safety boundaries, including process group termination, wall-time deadlines, and output byte caps to prevent resource exhaustion.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for downloading models and dependencies from Hugging Face and GitHub. All documented downloads target well-known repositories (e.g., unslothai/unsloth, ggml-org/llama.cpp) or the official Hugging Face Hub, which are considered safe sources.
[DATA_EXFILTRATION]: While the skill interacts with training data and model weights, the provided scripts (unsloth_preflight.py, decision_audit.py, specialization_audit.py) are strictly read-only and designed to prevent text leakage in logs. They emit only counts and codes, and the documentation explicitly warns against persisting credentials or transmitting private data to external providers.
[DYNAMIC_EXECUTION]: A call to eval() is present in verify_export.py for processing numeric token lists from shell output, and gpu_completion.cpp involves runtime compilation. These are constrained to specific, user-authorized verification tasks within isolated run directories rather than arbitrary input execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted JSONL datasets for training. To mitigate injection risks, the sft_audit.py and unsloth_preflight.py scripts implement strict schema validation, assistant-only loss masking, and tokenizer template verification to ensure instructions in the data do not override model behavior during the training process.

unsloth