README.md

metadata

configs:
  - config_name: models
    data_files:
      - split: models
        path: models.jsonl
  - config_name: benchmarks
    data_files:
      - split: benchmarks
        path: benchmarks.jsonl
tags:
  - benchmarks
  - evaluation
  - llm
  - leaderboard

LLM Benchmark Usage (2023–2026)

Which evaluation benchmarks 39 AI labs use to evaluate their models, and how that's changed over time — hand-built from 62 papers, technical reports, system cards, model cards, and blog posts, covering 128 models from 2023-07 to 2026-07.

from datasets import load_dataset

models = load_dataset("SaylorTwift/llm-benchmark-usage", "models")["models"]

models

One row per model, fully self-contained. 128 rows.

column	type	description
model_id	string	Hugging Face repo id for open-weight models (e.g. Qwen/Qwen3.5-397B-A17B), or a plain slug for closed API models (e.g. claude-opus-4-8, gpt-5.5, gemini-2.5-pro, grok-4)
lab	string	Organization/lab that released the model
release_date	timestamp	Release date. HF repo creation date for open-weight models (a proxy, not always the exact announcement date); hand-researched announcement/system-card date for closed models
source	string	Link to the paper/report/card/blog this model's evaluation suite was extracted from
benchmarks	list[string]	The evaluation suite: benchmark names, canonicalized (e.g. always GPQA-Diamond, never GPQA Diamond/GPQA-diamond). Values act as foreign keys into benchmarks.name where a metadata entry exists — the benchmarks split currently covers only a subset of all referenced benchmarks, so a name without a matching row there just means no metadata has been collected yet, not that the benchmark wasn't used. Per-source free-text categories were dropped when this field was flattened to plain names

Richer per-source metadata (paper type, title, notes, confidence flags) previously lived in a sources table alongside this one; it may move to a separate dataset later.

benchmarks

One row per benchmark, 125 rows: name, categories (list, e.g. agentic, math), modality (list), language (list), description, paper_url, implementation_url. models.benchmarks values reference this table by name; coverage is partial (125 of ~540 distinct referenced benchmarks) and will grow over time.

Benchmark name canonicalization

Benchmark names are deduplicated across ~250 raw name variants collected from primary sources (casing, hyphenation, and cross-lab transliteration differences — e.g. τ²-Bench / TAU2-Bench / TauBench V2 were all the same benchmark and are unified to TAU2-Bench). Genuinely distinct benchmark variants are kept separate on purpose (e.g. GPQA vs. GPQA-Diamond vs. GPQA Hard, or HumanEval vs. HumanEval+).

Known limitations

• Not a random or complete sample of all models ever released — built by snowballing outward from four 2026 HF leaderboards (GPQA, HLE, SWE-bench Pro, Terminal-Bench 2.0), a hand-picked set of closed models, and a curated flagship pull from ~18 additional labs added for historical depth back to 2023. Skews toward H1 2026.
• The Grok line (xAI) is sourced partly from screenshots of the original announcement pages (which block automated fetching) and partly from secondary aggregators where no screenshot was available.
• lab attribution is by HF namespace; community requantizations of another lab's checkpoint (e.g. exolabs/RedHat AI repacks of NVIDIA models) are attributed to the original lab, since requantizing isn't an independent benchmarking choice.
• Two models currently resolve to an empty benchmarks list due to pre-existing gaps in the source data: HelpingAI/Dhanishtha-2.0-0126 (its source's benchmark entries are all scoped to its 3 other models) and MiniMaxAI/MiniMax-H3 (its source has no benchmark entries recorded yet).