The Agent Skills Directory

[SAFE]: The skill serves as an educational repository for machine learning recipes. All logic related to training loops, model architectures (Transformers, CNNs), and optimization strategies (Muon, AdamW) follows established industry practices for development and research.
[REMOTE_CODE_EXECUTION]: A documentation snippet in references/domain-specific.md demonstrates checkpoint loading using torch.load(..., weights_only=False). While this is a standard legacy pattern in ML research, it allows for the deserialization of arbitrary Python objects and could theoretically lead to remote code execution if used with an untrusted checkpoint file. Modern best practices recommend using weights_only=True.
[EXTERNAL_DOWNLOADS]: The skill references multiple well-known third-party machine learning libraries and community-contributed CUDA kernels (such as flash-attn3). These references are used for legitimate technical documentation purposes within the ML ecosystem.

ml-training-recipes