The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data for model fine-tuning from varied sources including Hugging Face, S3, and local files. This creates a surface where maliciously crafted data could influence agent behavior during training or evaluation.
Ingestion points: Dataset loading via YAML configurations (e.g., path, data_files, pretraining_dataset) as described in references/other.md and references/dataset-formats.md.
Boundary markers: The skill relies on chat templates and Jinja2 formatting to delimit message turns (e.g., <|im_start|>/<|im_end|>), providing structural markers for processed content.
Capability inventory: The skill manages powerful execution capabilities such as cli.cloud.modal_.run_cmd for cloud commands (referenced in references/api.md), cli.train.do_train for training execution (references/api.md), and file system writes via convert.FileWriter (references/api.md).
Sanitization: While schema validation using Pydantic is mentioned for configurations, no explicit sanitization for user-provided dataset content is detailed in the reference documentation.
[COMMAND_EXECUTION]: The skill provides numerous examples and API functions for executing system commands. This includes the run_cmd function for Modal cloud integration, NCCL performance test commands, and the use of the torchrun launcher for distributed multi-node training.
[EXTERNAL_DOWNLOADS]: Installation and setup instructions involve downloading code and dependencies from external repositories. This includes fetching configurations from GitHub repositories and installing packages via PyPI using pip and uv.
[REMOTE_CODE_EXECUTION]: Documentation includes instructions for installing tools via remote scripts from well-known sources, such as the uv installer via a shell pipe from astral.sh, and installing specific package revisions directly from GitHub URLs.
[DYNAMIC_EXECUTION]: The skill makes extensive use of dynamic code techniques, including PyTorch's torch.compile for model acceleration and various monkey-patching modules (e.g., monkeypatch.lora_kernels, monkeypatch.gradient_checkpointing) to modify library and model behavior at runtime for performance optimizations.

axolotl