Mostly code-driven and actionable, but the '16× faster / 120h→7.5h' benchmark recurs at multiple points (When-to-use, GPU table, Performance benchmarks, references) and the Quick-start pipeline duplicates Stage 1/2 code; cost-comparison and star-count details pad the body.

The body is dominated by concrete, executable code with real import paths and specific parameters (e.g. WordCountFilter(min_words=50, max_words=100000), FuzzyDuplicates(num_hashes=260, num_buckets=20)), going well beyond pseudocode.

Stage 1→4 sequencing and the Common Crawl pipeline list are clear, but no validation checkpoints or error-recovery loops exist for these batch curation operations, capping the score at 2 per the batch-operation guideline.

Two real, well-signaled one-level references exist (filtering.md, deduplication.md, both verified present), but the ~370-line body keeps full multi-modal, classifier, benchmark, and cost content inline that could be split into additional reference files.

Lists multiple concrete actions — 'fuzzy deduplication (16× faster)', 'quality filtering (30+ heuristics)', 'semantic deduplication', 'PII redaction', 'NSFW detection' — rather than vague language.

Explicitly answers what it does (the feature list) and when to use it via the explicit 'Use for ...' trigger clause, satisfying both halves.

The 'Use for preparing high-quality training datasets, cleaning web data, or deduplicating large corpora' clause covers natural terms a user would say, alongside domain terms like 'LLM training' and 'deduplication'.

The GPU-accelerated LLM-training-data curation niche with RAPIDS and named alternatives (datatrove, dolma, Ray Data) makes it clearly distinguishable and unlikely to trigger for the wrong skill.