The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes patterns for executing system-level commands to facilitate training and benchmarking. Examples include the run_cmd function for Modal Cloud, the execution of NCCL performance tests (./build/all_reduce_perf), and the use of launchers like torchrun and accelerate to start training jobs.
[EXTERNAL_DOWNLOADS]: The framework allows for syncing resources and configuration from remote sources. Specifically, the cli.utils.fetch.fetch_from_github function downloads files from GitHub repositories, and cli.config.check_remote_config fetches and parses configuration files from HTTPS URLs.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic loading to support a wide range of model architectures and user-defined extensions. This is evidenced by the PluginManager which loads classes based on module strings and the monkeypatch.lora_kernels.get_attention_cls_from_config function which uses dynamic imports.
[INDIRECT_PROMPT_INJECTION]: As a tool for training on external data, the skill possesses an attack surface for indirect prompt injection via untrusted datasets. However, it provides robust documentation on mitigation strategies, including the use of Jinja2 chat templates, turn delimiters, and sequence masking with specific labels (e.g., -100) to ensure the model distinguishes between instructions and data.

axolotl