The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and analyze config.json files from the Hugging Face Hub to identify model architectures and implementation requirements.
Ingestion points: SKILL.md (Section 1.1) directs the agent to fetch model configurations from external repositories using tools like AutoConfig.from_pretrained.
Boundary markers: No specific delimiters or "ignore embedded instructions" markers are established for the ingested configuration data.
Capability inventory: The agent uses the ingested data to guide code implementation, registry updates (in _transformers/registry.py), and weight mapping logic across various scripts.
Sanitization: The instructions provide a specific list of architectural fields to extract (e.g., architectures, hidden_size, num_hidden_layers), which serves as a narrow schema for the ingested data.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of pre-trained model configurations and weights from the Hugging Face Hub, which is a well-known service for machine learning assets.
Source: Hugging Face Hub.
Method: Utilizes standard framework calls such as AutoConfig.from_pretrained and NeMoAutoModelForCausalLM.from_pretrained to retrieve model assets.

nemo-automodel-model-onboarding