The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation describes functionality for the axolotl fetch command and the check_remote_config utility, which are designed to download configuration files from remote HTTPS URLs and synchronize repository contents from the framework's official GitHub repository.
[DYNAMIC_EXECUTION]: The skill details Axolotl's plugin system, which involves dynamically importing and instantiating Python classes at runtime. It also documents the use of monkey-patching for applying optimizations like Flash Attention and Triton kernels to models during the loading and training process.
[INDIRECT_PROMPT_INJECTION]: The skill documents the processing of external datasets (e.g., JSONL, Parquet, and CSV) for fine-tuning, which represents an ingestion surface for untrusted data. * Ingestion points: Dataset loading utilities described in api.md, such as load_datasets and load_preference_datasets. * Boundary markers: The documentation describes the use of chat templates and delimiters (such as <|im_start|> and <|im_end|>) to separate conversation turns. * Capability inventory: Training (do_train), inference (do_inference), and model serving (do_vllm_serve) capabilities. * Sanitization: Not explicitly defined, though the skill outlines standard data preparation workflows using Hugging Face's datasets library.

axolotl