The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from web scrapes and S3 buckets for LLM training preparation.
Ingestion points: SKILL.md specifies reading Parquet files from common_crawl/ and s3://.
Boundary markers: No specific delimiters or 'ignore' instructions for the data content are mentioned.
Capability inventory: Performs file writing via to_parquet and executes GPU-accelerated processing.
Sanitization: Includes robust sanitization tools like PIIRedactor, NSFWClassifier, and QualityClassifier.
[COMMAND_EXECUTION]: Provides standard installation commands for the NeMo Curator package using uv pip in SKILL.md.
[EXTERNAL_DOWNLOADS]: Downloads pre-trained models from Hugging Face for quality classification and embedding, including models from nvidia, openai, and sentence-transformers.

nemo-curator