The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute management tasks such as verifying GPU status via nvidia-smi and running local training scripts (e.g., sft_train.py). These operations are necessary for the skill's purpose and are limited to the user's project environment.
[EXTERNAL_DOWNLOADS]: Automated installation of official machine learning libraries (unsloth, torch, transformers) is performed using uv or pip within a local virtual environment. This process facilitates the necessary setup for model training without affecting system-wide packages.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied training data in JSONL format.
Ingestion points: Data is read in scripts/sft_train.py and scripts/evaluate_model.py to facilitate fine-tuning operations.
Boundary markers: The scripts/data_validation.py module performs checks on JSON structure, required fields (ChatML/Alpaca), and duplicate records before any training occurs.
Capability inventory: The skill executes subprocesses for environment setup and writes model checkpoints/logs to the local disk.
Sanitization: While the raw content of training examples is not sanitized (as it is required for the learning objective), the strict schema validation prevents malformed data from interfering with the script's execution logic.

unsloth-workflows