The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation and scripts contain references to installing software via shell scripts piped from remote sources. Specifically, it mentions the installation of Ollama (ollama.com/install.sh) and Rust (sh.rustup.rs). These are well-known services used for local model serving and systems programming, and their use is standard in the machine learning development workflow.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of various external resources, including base model weights from HuggingFace, official Python packages from PyPI (e.g., unsloth, vllm, sglang, torch), and official Docker images (unsloth/unsloth). These components are standard for the described purpose of LLM training.
[COMMAND_EXECUTION]: The instructions provide a wide array of command-line operations for the user to execute, including Docker container management, model serving with vLLM and Ollama, and mobile deployment tasks using adb and ExecuTorch tools.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Untrusted data enters the training pipeline through load_dataset calls and the reading of local files like training_data.jsonl in reference/sft/sales_extractor_training.py and reference/grpo/basic_grpo.py.
Boundary markers: The skill employs system prompts that enforce structured output using XML tags (e.g., <reasoning>, <answer>) to help distinguish between model logic and generated data.
Capability inventory: The training environment has the capability to write large model files to disk, download remote model weights, and perform significant compute operations. It also provides instructions for serving the resulting models via external APIs.
Sanitization: The scripts utilize Unsloth's standard utilities such as get_chat_template and standardize_train_dataset to normalize data according to the model's expected format.

unsloth-training