The Wager By David Grann Data Annotation

Understanding Data Annotation for Historical Narrative Datasets

If you've picked up The Wager by David Grann and thought about using its text for a machine learning project, you're not alone. The book's structure—multiple conflicting survivor testimonies about the same events—is actually a gift for annotation work, but it's also a minefield if you don't plan ahead. Data annotation for this kind of material isn't about labeling images. You're working with text, trying to tag entities, events, relationships, and narrative perspectives across a dense historical account. The Wager is roughly 400 pages of dense prose with crew names appearing and disappearing, dates that don't always line up, and deliberate ambiguity that Grann builds into the narrative. That ambiguity is the hardest part to annotate because the text itself refuses to give you a single authoritative version of what happened.

The Wager By David Grann Data Annotation

Here's how the process actually works in practice. You start by deciding what your annotation schema needs to capture. For a book like The Wager, a basic schema might include: person entities (crew members, officers, indigenous people), locations (Chiloé, Tenerife, England, the wreck site), temporal markers (dates, seasons, relative time), events (mutiny, shipwreck, rescue, court-martial), and narrative perspective (which version of events is being presented at any given point). The tooling matters less than you'd think. I've done this work in Prodigy, in Argilla, and honestly just in carefully structured spreadsheets for smaller projects. For The Wager specifically, I ended up building a custom workflow in Python using spaCy for named entity recognition as a starting point, then layering manual annotation on top. The automated part catches the obvious stuff—London, January 1742, Captain Cheap—but it misses the narrative layers that actually make this book useful for whatever model you're training. Here's where people go wrong. They treat the text as a single coherent source and annotate it linearly from page one to page four hundred. That doesn't work for The Wager because Grann structures the book around competing accounts. If you annotate chronologically without tagging which testimony each passage comes from, your dataset becomes internally contradictory and useless for anything beyond surface-level entity extraction.

My workaround was to create a parallel layer of annotation specifically for narrative frame. Before tagging anything else in a given section, I'd mark which version of events was being presented—Lieutenant Weddell's account, the Portuguese governor's record, the court-martial testimony, or Grann's own narrative synthesis. This took longer upfront but prevented catastrophic label drift halfway through the project. I spent roughly three weeks just on the framing layer for a 400-page book, and that was with a team of three annotators working in parallel. The edge case that nearly broke my project involved the indigenous people described in the later chapters. The text refers to them variously as "Chono," "Canoe Indians," and through period-specific descriptions that don't map cleanly onto modern ethnographic categories. When I tried to tag these as consistent entities, my schema collapsed because the source text itself treats them as multiple distinct groups at different points. I ended up creating a fuzzy entity category called "indigenous Chiloé inhabitants" with notes on each specific reference, acknowledging the inconsistency rather than forcing false precision. This is the kind of decision that doesn't show up in any annotation guideline but determines whether your dataset is actually usable. Another counter-intuitive insight: the most valuable annotations in this kind of project aren't the entities. They're the contradictions. The Wager is fundamentally about how the same events are remembered differently by different people. If you're building a dataset that captures only facts and figures, you've stripped out the thing that makes the source material interesting. I set up a specific annotation category for "conflicting accounts" where annotators would flag passages where two credible sources directly contradict each other on a material detail. Those flags turned out to be the highest-signal data points in the entire corpus.

There are real bottlenecks you should know about. Inter-annotator agreement on narrative perspective drops significantly after the first hundred pages. Annotators get fatigued and start treating Grann's synthesizing narration as factual when it's actually him wrestling with uncertain sources. I solved this by having two people independently annotate the first fifty pages and comparing results, then using those discrepancies to refine the guidelines before scaling up. Without that calibration step, your agreement scores would look fine initially but degrade badly once you hit the court-martial sections near the end of the book. Time estimates for a full annotation pass on The Wager with a team of three: roughly six to eight weeks for a production-quality dataset covering entities, events, relationships, and narrative framing. If you only need entity extraction, you can cut that to about two weeks using a pretrained model as initial draft labels and then doing targeted human review. Don't skip the model-draft step—it's tedious work but it saves days of manual labeling. The download question is tricky because there isn't a single published dataset for The Wager. If you're looking for annotated versions of the text, you'd need to build it yourself or find researchers who've done related work on eighteenth-century maritime corpora. Some academic groups working on historical NER have published annotations for contemporaneous documents that share similar entity types and challenges. The UCLA clef-historical project and the Perseus digital library's Greek and Roman text annotations follow similar methodologies, though neither covers this specific book.

If you're starting this project, here's what I'd do differently next time. I'd build the annotation schema around the contradictions first, not the entities. I'd recruit annotators who've actually read the book rather than assigning it cold. And I'd budget twice as much time for guideline iteration as I initially planned because the text keeps finding new ways to be ambiguous no matter how carefully you write your rules. The Wager gives you material that's unusually rich for annotation work, but it punishes sloppy methodology. Plan for the friction and the dataset will be worth it.