VS: MaskLM vs Claude Haiku vs GLiNER
Pick a kind of document for a random example, then run all three on it with the same instructions. Each one is scored against a human-annotated answer key, and none of the examples were seen in training. Every round adds to the Frontier comparison.
Random example
Types
MaskLM
Baseline
Answer key (human annotation)
Where they differ
Claude Haiku answers with strings, which are located in the text. Scores follow the slider: “whole items right” matches each found item to one answer-key item; “text marked right” counts the characters marked.