The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The scripts/train_sft.py file includes a PEP 723 dependency on unsloth hosted at git+https://github.com/unslothai/unsloth.git. While essential for the skill's primary purpose, the unslothai organization is not in the trusted sources list, which could allow for the execution of unverified code if the repository were compromised.
COMMAND_EXECUTION (MEDIUM): The references/TROUBLESHOOTING.md file suggests executing sudo apt install nvidia-driver-535. Recommending high-privilege sudo commands for environment configuration poses a risk of privilege escalation if not handled carefully by the user.
INDIRECT_PROMPT_INJECTION (LOW): The skill processes untrusted external data during the training phase.
Ingestion points: DATASET_NAME (e.g., mlabonne/FineTome-100k) in scripts/train_sft.py.
Boundary markers: Absent; the training script uses standard SFTTrainer logic without explicit delimiters to prevent the model from learning malicious instructions embedded in the dataset.
Capability inventory: File system write access to outputs/, sub-process execution via huggingface-cli, and optional network uploads to Hugging Face and WandB.
Sanitization: None; the data is mapped directly into the training prompt via tokenizer.apply_chat_template.
CREDENTIALS_UNSAFE (LOW): SKILL.md suggests using export WANDB_API_KEY="your-key". While this is a placeholder and not a hardcoded secret, users should be reminded to use environment management tools or secret stores rather than plain text command history.

funsloth-hfjobs