The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documents utility functions like cli.utils.fetch.fetch_from_github and cli.config.check_remote_config which enable the retrieval of datasets, examples, and configurations from remote HTTPS and GitHub sources. It also references installation scripts for dependencies from well-known platforms like astral.sh.\n- [DYNAMIC_EXECUTION]: Axolotl uses a plugin system managed by integrations.base.PluginManager that dynamically loads Python modules based on user-provided configuration strings. It also employs runtime monkeypatching to optimize model components with custom Triton and Flash Attention kernels.\n- [COMMAND_EXECUTION]: The skill includes APIs for executing shell commands in cloud environments, specifically cli.cloud.modal_.run_cmd, which is used for managing training jobs on the Modal platform.\n- [INDIRECT_PROMPT_INJECTION]: As a tool for fine-tuning language models, the skill processes external dataset files, which could potentially contain malicious embedded instructions. This represents a known attack surface.\n
Ingestion points: Dataset file paths specified in the training configuration (e.g., JSONL, CSV, Parquet files).\n
Boundary markers: The documentation provides extensive instructions on implementing chat templates and delimiters (e.g., <|im_start|>) to ensure the model distinguishes between different conversation roles.\n
Capability inventory: The library provides broad capabilities including local and remote file system access, network operations for resource fetching, model hub interactions, and cloud-based command execution.\n
Sanitization: No specific content filtering or sanitization for prompt injection patterns is described; the system relies on structured parsing and proper template configuration.

axolotl