The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to install several third-party Python libraries including einops, rotary-embedding-torch, and flash-attn from public registries. It also provides commands to clone external GitHub repositories to access research implementations.
Evidence: git clone https://github.com/jquesnelle/yarn and https://github.com/ofirpress/attention_with_linear_biases references.
[COMMAND_EXECUTION]: The skill provides multiple shell commands for environment setup and model training, including script execution from cloned repositories.
Evidence: pip install commands in SKILL.md and python scripts/train.py in the fine-tuning section.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of large, untrusted external datasets (e.g., PG-19, arXiv, GitHub code) for model fine-tuning. This creates a vulnerability surface where malicious instructions embedded in the data could influence the agent during the fine-tuning or evaluation process.
Ingestion points: references/fine_tuning.md (e.g., load_dataset('pg19', ...)).
Boundary markers: None identified; data is processed as raw text for tokenization.
Capability inventory: The skill has capabilities for shell command execution (pip, git), file writing (model.save_pretrained), and network operations via HuggingFace API.
Sanitization: No specific sanitization or filtering logic is provided for the incoming dataset content.

long-context