I like the fact that there are various options for annotations, particularly when working with datasets which consist of conversations, documents, and unstructured texts. The workspace is tailored towards labeling rather than the spreadsheet way, and therefore it is easier to examine single examples and make sure that project guidelines are followed throughout the labeling process.

While we use Snorkel Flow to bulk-label large training datasets with Python rules, I personally use Datasaur Data Studio to curate and audit the Golden Test Suite for our insurance claims pipeline. When we ingest messy contents inventory spreadsheets, columns like 'Model#' and 'Brand or Manufacturer' are often left blank because adjusters type everything inside the raw 'Item Description' column. What I like best is how quickly I can switch between span-level entity tagging and row-level classification in the same spreadsheet-style view. Using keyboard hotkeys and LLM-assisted pre-labeling, I can highlight brands, model numbers, and dimensions buried inside item descriptions while also fixing low-confidence spans from our AssemblyAI audio transcripts. From there, I tag the item 'Condition', 'Item Age', and depreciation category so we have a clean ground-truth dataset to benchmark our Fireworks AI extraction models and Weights & Biases evaluation runs.

Datasaur makes data labeling easier with a simple interface, efficient annotation tools, and support for NLP projects.