The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents functions within the axolotl library that execute shell commands, specifically cli.cloud.modal_.run_cmd for cloud operations and cli.utils.train.launch_training for starting training runs.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading external configurations and resources through functions like cli.utils.fetch.fetch_from_github and cli.config.check_remote_config, which fetches and parses remote YAML or JSON files.
[DYNAMIC_EXECUTION]: The skill describes a plugin system that utilizes dynamic module loading (integrations.base.BasePlugin.load_plugin) and includes numerous modules for monkey-patching core libraries like PyTorch and Transformers to apply performance optimizations.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external datasets (e.g., JSONL, CSV, Parquet) from arbitrary local paths or remote URLs, creating a surface for indirect prompt injection. 1. Ingestion points: Data source paths and URLs defined in training configurations, as documented in references/dataset-formats.md. 2. Boundary markers: The documentation explains the implementation of chat template delimiters (e.g., <|im_start|>) to separate conversation turns. 3. Capability inventory: The skill facilitates file writing (FileWriter), shell command execution (run_cmd), and network access for data retrieval. 4. Sanitization: The library provides mechanisms to mask specific segments of input data (e.g., label: false in template-free formats) to ensure the model only learns from intended outputs.
[NO_CODE]: This skill consists entirely of documentation in Markdown format and does not ship with any executable scripts or binary files.

axolotl