The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process training data from external sources, creating an entry point for potential indirect prompt injection during the fine-tuning process.
Ingestion points: Processes datasets from local files, Hugging Face Hub, and remote storage like S3 or Google Cloud Storage.
Boundary markers: Utilizes chat templates and structural markers (e.g., ChatML tokens) to separate user instructions from assistant responses.
Capability inventory: Includes model training execution, file system access for saving checkpoints, and network access for fetching datasets.
Sanitization: Implements tokenization and formatting strategies to normalize raw data according to specified templates.
[COMMAND_EXECUTION]: The skill provides patterns for executing CLI tools necessary for ML workflows, such as axolotl for training, nvidia-smi for monitoring, and torchrun for distributed processing. It also includes functions to run commands on cloud infrastructure like Modal.
[REMOTE_CODE_EXECUTION]: The documentation includes instructions for downloading and executing setup scripts from well-known services (e.g., the uv installer from astral.sh) and provides utilities to fetch configuration files from the official project GitHub repository.
[DYNAMIC_EXECUTION]: To achieve high performance, the framework uses dynamic techniques such as monkey-patching of attention mechanisms and runtime integration of third-party plugins.

axolotl