Training data quality is the single most important factor in AI model performance. Data quality dimensions: 1) Accuracy: correct labels, factual content. 2) Completeness: sufficient examples per class/case. 3) Consistency: same label criteria applied uniformly. 4) Diversity: representative coverage of the target distribution. 5) Freshness: relevant to current conditions. Deduplication: near-duplicate text detection with MinHash LSH (faster than pairwise similarity), removes 20-40% of common web-scraped datasets that contain duplicates affecting generalization. Tools: dedup library, minhash from datasketch. Annotation quality control: 1) Inter-annotator agreement (IAA): Cohen Kappa or Krippendorff alpha. Good agreement: kappa > 0.7. If below 0.6, revisit annotation guidelines. 2) Annotation guidelines: document edge cases, provide examples for each class, include anti-examples (what is NOT this class). 3) Gold standard test items: inject known-correct examples throughout annotation batches to measure annotator accuracy, remove low-quality annotators. 4) Multiple annotations: critical labels annotated by 3+ annotators, use majority vote or expert adjudication for disagreements. Data governance: data lineage tracking (source, version, license), PII scanning and removal before training, documentation of known limitations and biases. Datasheet template: dataset motivation, composition, collection process, preprocessing, uses, distribution, maintenance.