TOON

lastmod 2025-11-19created 2025-11-191 edits

A token-efficient data serialization format for LLM input

About

• Token-Oriented Object Notation
• A serialization format designed to reduce token usage when feeding data to an LLM
  • Represents JSON data losslessly while achieving both token efficiency and readability
• Structure optimized for uniform arrays of objects; removes repeated keys for 30–60% token savings
  • A hybrid format combining YAML's indentation-based structure with CSV's tabular layout
  • For deeply nested or non-uniform data, JSON is more efficient
• Includes explicit structural information so LLMs can reliably parse and validate the data
  • Declares array length ([N]) and field definitions ({id,name,role}) to improve structure recognition
• Minimal syntax design with a concise grammar that drops unnecessary brackets, quotes, and punctuation
  • Indentation-based hierarchy; supports CSV-like row-by-row data streams
• Bidirectional JSON ↔ TOON conversion via a CLI tool
  • Options include automatic format detection, delimiter selection (,, \t, |), and token savings statistics output
• API structure
  • encode() serializes JSON, decode() deserializes
  • Options for indentation, delimiter, and length marker (#)
• Benchmarks show an average of 21–60% token savings over JSON and 73.9% LLM query accuracy
  • Slightly larger than CSV, but structural validation improves LLM reliability
• Format rules
  • Strings are quoted only when needed; automatically quoted if they contain the delimiter
  • Numbers, booleans, dates, etc. are converted to LLM-friendly forms
• Implementations in many languages
  • Official: Python, Rust (in development)
  • Community: Go, Java, Swift, C++, .NET, Ruby, etc.

See also

• LLM

Favorite site

• Github - toon-format/toon
• TOON - A token-efficient data serialization format for LLM input | GeekNews