The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the Tiny Shakespeare dataset from a public GitHub repository for use in character-level training demonstrations. Evidence: Source URL https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt is referenced in references/data.md and references/training.md.- [DYNAMIC_EXECUTION]: The training script utilizes torch.compile() to optimize the model graph into efficient kernels at runtime. Additionally, it uses the pickle and torch.load modules for serializing and deserializing local metadata and model checkpoints. Evidence: Found in the training loop implementation within references/training.md.- [INDIRECT_PROMPT_INJECTION]: The skill processes large text datasets from external sources, which is a standard attack surface for language model training where malicious instructions could be embedded in the training corpus. 1. Ingestion points: Dataset downloads from GitHub and HuggingFace integration in references/data.md. 2. Boundary markers: Not present. 3. Capability inventory: File system writes (.bin, .pkl), network requests for data retrieval, and GPU-accelerated code execution. 4. Sanitization: No sanitization is applied to the raw training text.- [COMMAND_EXECUTION]: The skill provides detailed shell commands for environment setup, data preparation, and distributed model training. Evidence: Instructions for pip install, python prepare.py, and torchrun are provided in SKILL.md and references/training.md.

nanogpt