AllChrome ExtensionEssayGamesGaussian SplatsMLOpen SourcePatentPlaygroundQuantified SelfSelf-HostedWritingArchived

OCR on Patent Figures with DeepSeek-OCR

12 tests extracting text and reference numbers from patent figure sheets using DeepSeek-OCR, GLM-OCR, and rotation heuristics.

PatentMLWritingEssay

Maps of Osaka

Osaka in the maps its own library keeps, in order: a coastline nobody had seen, a castle buried under another castle, a map that took four years to cut after it was licensed, and the canals that are now an elevated road.

WritingEssay

Failing to Train DeBERTa to Detect Patent Antecedent Basis Errors

Fine-tuned DeBERTa to detect antecedent basis errors in patent claims. 90% F1 on synthetic data, 8% recall on real data. A failure—but a diagnostic one.

PatentMLWritingEssay