The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process massive datasets from external sources like Common Crawl, which are untrusted and may contain adversarial content. This creates an indirect prompt injection surface where malicious data could attempt to influence the agent's logic or the resulting model's behavior.
Ingestion points: Data is loaded from external files via DocumentDataset.read_parquet and DocumentDataset.from_pandas as shown in SKILL.md.
Boundary markers: The provided documentation and examples do not define specific delimiters or instructions for the agent to isolate untrusted data from processing logic.
Capability inventory: The skill uses RAPIDS and NeMo Curator for data transformation; no evidence of arbitrary command execution or unauthorized network exfiltration was found in the scripts.
Sanitization: The skill includes dedicated modules for PIIRedactor and NSFWClassifier to identify and remove sensitive or harmful content from processed data.
[EXTERNAL_DOWNLOADS]: The skill initiates the installation of specialized GPU-accelerated libraries including nemo-curator, cudf, dask, and rapids. These are well-known high-performance data processing tools provided by the NVIDIA and RAPIDS ecosystem.

nemo-curator