LLM-Efficient Adaptive Notation — a token-optimized serialization format for structured data in LLM context windows.
Feeding structured data to LLMs wastes tokens on JSON syntax: repeated keys, quotes, braces, commas. LEAN eliminates this overhead while maintaining lossless round-trip fidelity.
Benchmark results (avg token savings vs JSON compact, 12 datasets):
| Format | Avg Savings | Round-trip |
|---|---|---|
| LEAN | -48.7% | ALL PASS |
| ZON | -47.8% | ALL PASS |
| TOON | -40.1% | ALL PASS |
| ASON | -39.3% | SOME FAIL |
npm install lean-formatimport { encode, decode } from 'lean-format';
const data = {
users: [
{ id: 1, name: "Alice", active: true },
{ id: 2, name: "Bob", active: false }
],
meta: { version: "2.1.0", debug: false }
};
const lean = encode(data);
const back = decode(lean);
// back deep-equals dataJSON (109 tokens):
{"users":[{"id":1,"name":"Alice","active":true},{"id":2,"name":"Bob","active":false}],"meta":{"version":"2.1.0","debug":false}}LEAN (46 tokens):
users[2]:id name active
1 Alice T
2 Bob F
meta.version:2.1.0
meta.debug:F
- T/F/_ for true/false/null (single char vs 4-5)
- No space after colon in key:value pairs
- Tab delimiter (most token-efficient separator)
- Tabular arrays — column headers declared once, data in rows
- Semi-tabular — non-uniform object arrays factor shared keys into columns, extra keys inline as
key:value - Hybrid dot-flatten/block — tries both, picks shorter
- Bare strings — no quotes unless ambiguous
Encodes a JSON-compatible value to LEAN format. Throws on invalid keys (dots, spaces, colons), NaN, or Infinity.
Decodes LEAN text back to a JSON-compatible value. Throws LeanParseError on invalid input.
Error class with line and content properties for debugging parse failures.
See SPEC.md for the full format specification.
MIT
