The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the Tiny Shakespeare dataset from a remote repository during data preparation. This is a standard procedure for retrieving training data from common developer resources.
Evidence: https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt is fetched using the requests library in references/data.md and references/training.md.
[EXTERNAL_DOWNLOADS]: The skill uses established machine learning libraries to download large-scale datasets and pre-trained model weights from official repositories.
Evidence: Uses the datasets library to load openwebtext and the transformers library to fetch gpt2 model checkpoints from HuggingFace.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of arbitrary text files for model training, which represents a potential surface for indirect prompt injection if the source data is malicious.
Ingestion points: The skill processes user-supplied text files (e.g., my_data.txt) as described in SKILL.md (Workflow 4) and references/data.md.
Boundary markers: There are no explicit delimiters or warnings to ignore instructions embedded within the training text.
Capability inventory: The skill uses subprocess.run or direct shell execution to launch training and sampling scripts based on the processed data.
Sanitization: No sanitization or filtering is applied to the raw text content before tokenization and training.

nanogpt