The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements the trust_remote_code=True parameter within the setup_qlora_model function for AutoModelForCausalLM.from_pretrained. This configuration allows the AI agent to execute custom code defined in a model's repository on the Hugging Face Hub. If used with an untrusted or compromised model ID, this could lead to arbitrary code execution on the host system.
[EXTERNAL_DOWNLOADS]: The skill utilizes the transformers and datasets libraries to download model weights and training data from the Hugging Face Hub. While the provided examples target reputable organizations (Meta and Databricks), the code is designed to fetch and process assets from any remote repository.
[COMMAND_EXECUTION]: The skill provides instructions for initiating distributed training using the accelerate launch command through the Bash tool. This involves executing system-level processes to manage multi-GPU training environments.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing external datasets through the load_dataset function.
Ingestion points: Data enters the context via load_dataset in the data preparation section of SKILL.md.
Boundary markers: Absent; the skill does not use delimiters or instructions to ignore embedded commands within the dataset.
Capability inventory: The skill possesses the Bash tool for command execution and WebFetch for network operations.
Sanitization: Absent; data from the dataset is directly interpolated into a string template using the format_instruction function.

moai-ml-llm-fine-tuning