Frontier comparison

How MaskLM compares with a frontier LLM (Claude Haiku 5.5) and a popular open-source extractor (GLiNER) on documents none of them trained on, each scored against a human-annotated answer key. Every round played in VS mode or on the Redaction page adds to these numbers.

Score by kind of document

MaskLMClaude Haiku 5.5GLiNER

Scoreboard

loading…

Whole items right (set F1): each item found must match one answer-key item of the same type. Text marked right (coverage F1): the right characters are marked, however they are grouped. MaskLM is scored at its calibrated threshold for the task, GLiNER at 0.5, Claude Haiku as it answered (its strings located in the text). Play more rounds →

Masked-RAG: answering from a compressed context

loading…