SaylorTwift/llm-benchmark-usage

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.

Load either table with load_dataset, selecting the config by name:

from datasets import load_dataset

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

models

One row per model. 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_id	string	Foreign key into the sources table — which paper/report/card describes this model's evaluation
leaderboards	list[string]	Which HF Hub leaderboards (if any) this model appears on, e.g. ["hle", "SWE-bench_Pro"]. Empty for closed API models, which aren't on any HF leaderboard

sources

One row per paper/report/card/blog. 62 rows. Several models can share one source (e.g. one paper covering a whole model family).

column	type	description
id	string	Primary key, matches models.source_id
type	string	One of paper, report, model_card, blog
title	string	Title of the paper/report/card/post
url	string, nullable	Link to the source
arxiv_id	string, nullable	arXiv ID if this is an arXiv paper
models	list[string]	Every model_id this source's benchmark list applies to by default
notes	string, nullable	Caveats — e.g. data-quality flags, which benchmarks are model-specific vs. shared, confidence level
benchmarks	list[struct]	The evaluation suite. Each item has name (benchmark name, canonicalized — e.g. always GPQA-Diamond, never GPQA Diamond/GPQA-diamond), category (free-text category as written in the source, e.g. knowledge, agentic_coding, preparedness_bio_chem), and models (nullable list[string] — when null, this benchmark applies to every model in the source's models list; when set, it applies only to the listed models, because some sources cover multiple models that weren't all evaluated identically, e.g. a base vs. instruct pair, or different size tiers of the same family)

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 — see each Grok source's notes field for exactly which benchmarks are verified vs. secondary-sourced.
• 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.