The world's first fine-grained writing preference dataset created from inline accept/reject decisions inside real documents.

TypeOS captures micro-edits (word choice, tone, structure, formatting, semantics), builds taste vectors, updates reward models, and enables personalized writing models. Unlike traditional RLHF which relies on binary "thumbs up/down" on entire responses, our approach leverages the granular edits users make to their own documents.

Micro-Level Preference Signals

Each edit generates micro-level preference supervision. No other platform collects real writing edits at this resolution. This becomes a new category of RLHF: Inline Revision Feedback (IRF).

Personalized Taste Vectors

By aggregating thousands of these micro-decisions, TypeOS constructs a high-dimensional "Taste Vector" for each user or organization. This vector informs the model about preferred:

• Tone: Formal vs. Casual, Direct vs. Diplomatic
• Structure: Dense paragraphs vs. Bulleted lists
• Vocabulary: Simple vs. Academic
• Formatting: Oxford commas, capitalization rules, spacing

Core Metrics