The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation and training scripts suggest installing dependencies from official, well-known sources. Evidence includes instructions for installing Ollama (curl -fsSL https://ollama.com/install.sh | sh) and Rust (https://sh.rustup.rs). The scripts also include commands to install the Unsloth framework directly from its official GitHub repository (git+https://github.com/unslothai/unsloth.git).\n- [COMMAND_EXECUTION]: The skill instructs the agent to execute various shell commands as part of the model development lifecycle. This includes running Docker containers (docker run), managing local LLM instances via Ollama (ollama create), and performing model quantization/export tasks.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external datasets for model training, which represents a potential surface for indirect prompt injection. \n
Ingestion points: Data is loaded using datasets.load_dataset in reference/grpo/basic_grpo.py and reference/sft/sales_extractor_training.py.\n
Boundary markers: The training scripts do not implement specific boundary markers or 'ignore' instructions for the training data, which is standard for machine learning workflows.\n
Capability inventory: The skill has capabilities for file system writes (checkpoints, exported models, telemetry) and network access (downloading base models and datasets).\n
Sanitization: No input sanitization is performed on the training data as the primary purpose is for the model to learn from the provided examples.

unsloth-training