The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and downloads pre-trained models and datasets from well-known and official repositories.
Fetches quality classifiers and speech-to-text models from NVIDIA's official Hugging Face repository (nvidia/quality-classifier-deberta, nvidia/stt_en_fastconformer_hybrid_large_pc).
Downloads embedding models from the Sentence Transformers organization (sentence-transformers/all-MiniLM-L6-v2).
Accesses the CLIP model from OpenAI's official repository (openai/clip-vit-base-patch32).
[INDIRECT_PROMPT_INJECTION]: The skill defines a data curation pipeline that ingests data from untrusted external sources, such as web scrapes (Common Crawl).
Ingestion points: Processes external data via DocumentDataset.read_parquet("common_crawl/*.parquet") in SKILL.md.
Boundary markers: No explicit instruction-isolation markers are used in the provided code snippets.
Capability inventory: The skill performs file system writes using .to_parquet() to store curated results.
Sanitization: Implements robust sanitization through 30+ quality filters, a PIIRedactor for sensitive information removal, and an NSFWClassifier to filter toxic content.

nemo-curator