Skip to content
DevMeme
6779 of 7435
TOON Format Beats JSON for LLM Data Retrieval Accuracy While Using Fewer Tokens
AI ML Post #7433, on Nov 18, 2025 in TG

TOON Format Beats JSON for LLM Data Retrieval Accuracy While Using Fewer Tokens

Description

A dark-themed benchmark chart titled 'Per-Model Accuracy' showing 'Accuracy across 4 LLMs on 209 data retrieval questions'. Four models are compared: claude-haiku-4-5-20251001, gemini-2.5-flash, gpt-5-nano, and grok-4-fast-non-reasoning. Each model shows horizontal bar charts comparing six data formats: TOON, JSON, YAML, XML, JSON compact, and CSV. TOON consistently ranks #1 or near-top across all models. For claude-haiku: TOON 59.8% vs JSON 57.4%. For gemini-2.5-flash: TOON 87.6% vs JSON 77.0%. For gpt-5-nano: TOON 90.9% vs JSON 89.0%. For grok-4-fast: TOON 57.4% vs JSON 55.5%. The bottom states the key tradeoff: 'TOON achieves 73.9% accuracy (vs JSON's 69.7%) while using 39.6% fewer tokens on these datasets.'

Categories

AI ML DataFormats Performance

Common Tags

Llm Json DataFormats Yaml PerformanceOptimization BenchmarkingTools Xml CostOptimization Csv

Comments

21
Anonymous ★ Top Pick We spent a decade declaring 'JSON won the format wars' only for LLMs to show up and say 'actually, this format you never heard of parses better and costs 40% less -- you're welcome.'
  1. Anonymous ★ Top Pick

    We spent a decade declaring 'JSON won the format wars' only for LLMs to show up and say 'actually, this format you never heard of parses better and costs 40% less -- you're welcome.'

  2. Anonymous

    A new data format that's more accurate and uses 40% fewer tokens? I'm sure the benchmarks were run on a perfectly curated dataset that only contains the word 'aardvark'

  3. @la_gerva_stock 7mo

    What part of it supposed to be a meme material?

  4. @Hollow_Arigo 7mo

    Can someone explain this one?

  5. @Broken_Cloud_1 7mo

    What

  6. @mrYakov 7mo

    Did this benchmark just test llm ability to read info from some formats ? Like, you give llm json and query like data["data"][0]["data"] ?

    1. @Mikle_Bond 7mo

      Plain-English questions about data. They have examples here: https://github.com/toon-format/toon?tab=readme-ov-file#question-types

  7. @q_rsqrt 7mo

    ai bros be benchmarking that TROON

    1. @NickNirus 7mo

      troonin my model rn

      1. @q_rsqrt 7mo

        ngmghhhhh

  8. @Mikle_Bond 7mo

    So, basically, it is Yaml2. Nice.

  9. @vgy4sw 7mo

    Who even chooses those names bruh 💔🥀

  10. Егор 7mo

    https://www.improvingagents.com/blog/toon-benchmarks

  11. @deadgnom32 7mo

    toon toon toon toon toon sachoor

    1. @vgy4sw 7mo

      https://t.me/devs_chat/176551 brotha

  12. @Mikle_Bond 7mo

    I think Yaml could do the same with a bit of type-awareness hikes: !!csv id,name,distanceKm,elevationGain,companion,wasSunny 1,Blue Lake Trail,7.5,320,ana,true 2,Ridge Overlook,9.2,540,luis,false 3,Wildflower Loop,5.1,180,sam,true

  13. @vgy4sw 7mo

    https://t.me/devs_chat/176551

  14. @mordegaard 7mo

    TOON results cannot be consistent enough because modern LLMs weren't trained on such data. So the results will differ regarding on how good did you describe the format in your prompt. The model can easily hallucinate out and start producing YAML response because TOON and YAML are quite familiar

  15. @vladfaust 7mo

    Stupid hype retards, my X feed is plagued with this shit. Also weird-angled Frieren. 😭

  16. @qtsmolcat 7mo

    Fun fact: most benchmarks are meaningless because if you alter the question even a bit the LLM flops

    1. @DerKnerd 7mo

      probably even for the benchmarks made between different models

Related Memes

Museum Of Meaningless Metrics

Museum Of Meaningless Metrics

DeveloperProductivity
Le Chaton Fat Wins Benchmarks

Le Chaton Fat Wins Benchmarks

AI ML
The Longest Government Tool Call

The Longest Government Tool Call

AI ML
Use J and K for navigation