The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches the Tiny Shakespeare dataset from a public repository on GitHub (raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt) using the requests library within references/data.md. This is used to provide training data for the educational examples.
[EXTERNAL_DOWNLOADS]: The data preparation scripts in references/data.md and references/training.md utilize the HuggingFace datasets library to load the openwebtext and wikipedia datasets, which are standard resources in the machine learning community.
[DYNAMIC_EXECUTION]: The training logic in references/training.md employs torch.compile(model), a PyTorch 2.0 feature that performs runtime optimization of the model graph. This is a performance-enhancing feature and does not involve executing untrusted code or scripts from external sources.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data, as it is designed to train models on custom user-provided text files or external web datasets. However, since the data is converted into numerical tokens for tensor-based training, the risk of instruction override or execution is minimal in this context.

nanogpt