The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions in SKILL.md and references/fine_tuning.md to clone an external GitHub repository (github.com/jquesnelle/yarn) and execute a training script (scripts/train.py) within it. This represents the execution of remote code from an unverified source.
[EXTERNAL_DOWNLOADS]: References several external repositories for specific implementations, including implementations for ALiBi (github.com/ofirpress/attention_with_linear_biases) and rotary embeddings (github.com/lucidrains/rotary-embedding-torch). It also instructs the agent to download large-scale datasets from the HuggingFace Hub.
[COMMAND_EXECUTION]: Contains numerous shell commands for installing Python dependencies and running local scripts for model fine-tuning, evaluation, and deployment.
[DATA_EXFILTRATION]: Provides instructions to use HfApi().upload_folder to upload models to the HuggingFace Hub, which involves network transmission of local file contents to an external service.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes long documents from external datasets (pg19, arXiv) and interpolates them into training or evaluation pipelines.
Ingestion points: datasets.load_dataset calls in references/fine_tuning.md and references/rope.md.
Boundary markers: Absent; there are no delimiters or warnings to ignore instructions within the ingested data.
Capability inventory: Shell execution via python scripts/train.py, file writes via model.save_pretrained, and network operations via HfApi.upload_folder.
Sanitization: Absent; external data is processed directly without filtering or validation.

long-context