135,624 comparisons among 53 agents.
| Description | Human preference votes from LMArena's text arena, 2025-04-17 to 2025-07-24: 135,624 votes in which a person saw two anonymous language models answer their own prompt and picked one, called a tie, or said both were bad. 53 named models are the contestants, 115,363 evaluation sessions are the raters. agent_a wins 48,559, agent_b 49,782, and 37,283 votes are draws (21,529 ties and 15,754 both_bad; `outcome` keeps the distinction that `winner` collapses). homefield is "agent_b" in every row, and that is measured rather than assumed: slot assignment is balanced (each model is agent_a in 50.00% of its battles), yet agent_a wins only 49.38% of the 98,341 decided battles (z = -3.90), the same figure within matchups. The effect is real but small -- carry a slot term in an Elo or Bradley-Terry fit, do not read it as a home advantage. `rater` is the evaluation session re-keyed to an integer, with `eval_order` giving the vote's place in that sitting (1 to 28); it is a weak person identifier, since 102,166 of the sessions hold a single vote and nothing links two sessions by one person. Also kept, all from LMArena's own per-vote annotations: language (126 values, en 71,169, pl 13,813, und 11,833), turns, is_code, math, creative_writing, instruction_following, and hardness, the count of seven prompt criteria met (0 to 7, skewed high: 30,644 votes score 7). The prompts, the model answers and the style-control token counts are not taken. 10 source rows were dropped for drawing the same model on both sides; nothing else. |
|---|---|
| Reference | LMArena. (2025). arena-human-preference-140k [Data set]. Hugging Face. https://huggingface.co/datasets/lmarena-ai/arena-human-preference-140k |
| Licence | CC BY 4.0 |
| Source data | https://huggingface.co/datasets/lmarena-ai/arena-human-preference-140k |
| Comparisons | 135,624 |
|---|---|
| Agents | 53 |
agent_aagent_bcreative_writingdateeval_orderhardnesshomefieldinstruction_followingis_codelanguagemathoutcomeraterscore_ascore_bturnswinner
Or load it directly in R (no account needed with irw 1.4.0 or later) or Python (needs a free Redivis account; the first call opens a browser to sign in):
# R
install.packages("irw")
library(irw)
df <- irw_fetch("lmarena_text_2025", source = "comp")
# Python
pip install irw
import irw
df = irw.fetch("lmarena_text_2025", source="comp")
Or in Stata, which reads the same no-account CSV as the button above:
* Stata
net install irw, from("https://raw.githubusercontent.com/itemresponsewarehouse/Stata-pkg/main/")
irw_fetch lmarena_text_2025, clear source(comp)
@dataset{lmarena2025arena,
title={arena-human-preference-140k},
author={LMArena},
year={2025},
publisher={Hugging Face},
url={https://huggingface.co/datasets/lmarena-ai/arena-human-preference-140k}
}
| IRW version | v712 |
|---|---|
| Redivis dataset | irw_competitions v7.1 |
| Redivis dataset DOI | 10.57761/3r1w-t660 |
| Manifest pin for this IRW version | v7.1 |
| Metadata source | irw_meta v68.0 |