The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides a dataset preparation script and training workflows that ingest untrusted external data in CSV, JSON, and JSONL formats.
Ingestion points: The load_input function in scripts/prepare_dataset.py and the training examples in references/fine-tune-guide.md ingest external files.
Boundary markers: No explicit boundary markers or "ignore embedded instructions" warnings are implemented when the agent processes these datasets.
Capability inventory: The skill utilizes file system access (Read/Write) and shell command execution (Bash), providing a surface that could be targeted by malicious instructions embedded in training data.
Sanitization: No sanitization or validation of dataset content is performed prior to processing or formatting.
[EXTERNAL_DOWNLOADS]: The skill references and interacts with well-known machine learning services and repositories.
Fetches model weights and configurations from Hugging Face's official model hub (e.g., 'mistralai/Mistral-7B-v0.1').
References standard, reputable machine learning libraries including transformers, peft, and unsloth.
[COMMAND_EXECUTION]: The skill facilitates the execution of dataset preparation and training tasks using Python.
Utilizes uv run to execute the local scripts/prepare_dataset.py and runs inline Python training loops via python3 -c for model optimization tasks.

fine-tune