Encoder models score every label in one pass, no decoding loop

Liquid AI's classification example assigns a document to several categories by calculating a separate score for every label in one model pass. It fine-tunes the company's small encoders for this task, avoiding the need to generate and then interpret a written answer. Announcement.

The linked cookbook example explains the decisions needed around those scores: train on labeled documents, select score thresholds using validation data, then evaluate on held-out examples. More than one label can clear its threshold. This turns the general-purpose encoder into a specialized categorization component with directly usable outputs.