The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes various system and Python commands to manage the training environment.
Uses nvidia-smi to detect GPU capabilities and driver versions.
Utilizes uv or pip to create a virtual environment and install necessary Python packages.
Spawns subprocesses to run internal workflow scripts like sft_train.py, evaluate_model.py, and export_model.py.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of external resources required for machine learning.
Installs the unsloth library and its dependencies from official package registries.
Downloads pre-trained models and tokenizers from Hugging Face based on user configuration.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data which could potentially contain malicious instructions.
Ingestion points: Reads user-supplied dataset files in JSONL format via scripts/data_validation.py and scripts/sft_train.py.
Boundary markers: Implements strict data validation checks to ensure records match expected ChatML or Alpaca formats before processing.
Capability inventory: Possesses the capability to execute training and evaluation scripts which involve substantial GPU computation and file system writes.
Sanitization: Employs a dedicated configuration validator (scripts/workflow_config.py) to enforce schema constraints and prevent unexpected execution parameters.

unsloth-workflows