The Agent Skills Directory

EXTERNAL_DOWNLOADS (LOW): The training script scripts/train_sft.py defines a remote dependency on the unsloth library via GitHub rather than a standard package registry.
Evidence: unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git in the PEP 723 dependency block.
COMMAND_EXECUTION (LOW): The troubleshooting guide suggests using elevated privileges to install GPU drivers.
Evidence: sudo apt install nvidia-driver-535 in references/TROUBLESHOOTING.md.
CREDENTIALS_UNSAFE (LOW): The skill instructions facilitate the injection of Hugging Face and Weights & Biases secrets into remote environments.
Evidence: The environment variable configuration in SKILL.md (Step 4) for HF_TOKEN and WANDB_API_KEY.
PROMPT_INJECTION (LOW): The skill possesses a surface for indirect prompt injection by ingesting external datasets for training.
Ingestion points: DATASET_NAME = "mlabonne/FineTome-100k" in scripts/train_sft.py.
Boundary markers: Absent; uses standard chat template formatting.
Capability inventory: The skill performs model weights saving and potential local script execution via standard ML libraries.
Sanitization: Absent; the skill relies on the inherent safety of the upstream dataset.

funsloth-runpod