Skip to main content

Benchmark Summary Example

What this example shows​

This example demonstrates SDIF describing itself — a BenchmarkSummary document that records the results of a token-efficiency comparison across data interchange formats.

Key concepts covered:

  • SDIF is self-hosting: the same format used to exchange structured data is used to record benchmark results
  • Scalar fields capture run metadata (run_at, corpus, host) without nesting
  • A results table with five columns delivers a dense, machine-readable summary across formats
  • Relations record the tokenizer used and the corpus that was measured, keeping provenance auditable
  • The SDIF AI projection shows how the same information compresses further for model context windows

The semantic_density column is the ratio of information content to token count — the core metric tracked by the SDIF benchmark suite.

SDIF​

@sdif 1.0
@profile source
kind BenchmarkSummary
id benchmark.token_efficiency.20260520
schema sdif.benchmark_summary.v1
authority Canonical
lifecycle Active
run_at 2026-05-20T09:00:00Z
corpus sdif.corpus.representative.v1
host ci.runner.linux.x64

results[format,bytes,tokens,semantic_density,roundtrip]:
  sdif	1420	312	0.91	lossless
  sdif.ai	980	198	0.94	lossless
  json	3210	748	0.62	lossless
  yaml	2870	681	0.65	lossless
  toon	2150	510	0.70	lossless

rel:
  benchmark.token_efficiency.20260520 tokenized_by cl100k_base
  benchmark.token_efficiency.20260520 measured_on sdif.corpus.representative.v1

SDIF AI​

The AI projection drops @profile, provenance fields, and the schema declaration. Relations are grouped by subject using the rel[subject]: syntax.

@sdif.ai 1.0
kind BenchmarkSummary
id benchmark.token_efficiency.20260520
run_at 2026-05-20T09:00:00Z

results[format,bytes,tokens,semantic_density,roundtrip]:
  sdif	1420	312	0.91	lossless
  sdif.ai	980	198	0.94	lossless
  json	3210	748	0.62	lossless
  yaml	2870	681	0.65	lossless
  toon	2150	510	0.70	lossless

rel[benchmark.token_efficiency.20260520]:
  tokenized_by cl100k_base
  measured_on sdif.corpus.representative.v1

Notes​

SDIF describing itself

The BenchmarkSummary document is a natural fit for SDIF's own benchmark results. The results table is wide enough to be dense but narrow enough to remain readable. A JSON equivalent of the same document would use roughly 2.3x the tokens — a ratio visible in the results table itself.

results table

Five columns capture everything needed to reproduce or dispute a benchmark finding:

  • format — the interchange format under test
  • bytes — raw byte size of the representative payload in that format
  • tokens — token count using the tokenizer named in rel:
  • semantic_density — information-per-token ratio (higher is better); computed as the ratio of semantic content units to token count relative to the sdif baseline
  • roundtrip — lossless if the format can represent and recover all semantic content; lossy if it cannot

Relations

Two predicates capture provenance without cluttering the scalar fields:

  • tokenized_by — names the tokenizer used (e.g. cl100k_base for GPT-4-class models); results change if the tokenizer changes, so this is load-bearing
  • measured_on — references the corpus identifier; the corpus defines what payload was used for each format, making results reproducible

sdif.ai row

The AI projection appears as its own row in results. This lets the benchmark answer the question "how much does the AI projection save?" without a separate document. At 0.94 semantic density and 198 tokens, sdif.ai is the densest representation in the table.

Try it locally​

# Parse the benchmark summary
sdif parse benchmark.sdif

# Convert to JSON for downstream tooling
sdif to-json benchmark.sdif

# Generate the AI projection from the source document
sdif ai benchmark.sdif

# Validate the summary against its schema
sdif validate benchmark.sdif --schema schema.sdif