lmarena_text_55k

57,477 comparisons among 64 agents.

About this table

DescriptionLMArena text-arena preference votes, the 55k release LMArena published for the Kaggle competition "LMSYS - Chatbot Arena Human Preference Predictions". 57,477 votes between 64 named language models; a person saw two anonymous models answer their own prompt and picked one or called it a tie. agent_a wins 20,064, agent_b 19,652, 17,761 draws. Two warnings. First, this release has no both_bad category, so its "tie" silently pools genuine ties with the both-bad votes the live arena recorded separately -- do not compare its 30.9% draw rate with lmarena_text_2025's without that in mind, and there is no `outcome` column here for the same reason. Second, there are no timestamps at all, so the table cannot be ordered or sliced by time, and it carries no date column. No homefield column either: slot assignment is balanced and agent_a wins 50.52% of the 39,716 decided votes, z = +2.07, short of the z = -3.90 measured on lmarena_text_2025. Distinct from lmarena_text_2025 despite the shared source organisation: the id schemes are incompatible (ten-digit integers here, UUIDs there) so the two cannot be joined, but this release predates the 2025 window and its 64-model roster still includes vicuna, koala, alpaca and llama-2.
ReferenceLMArena. (2024). arena-human-preference-55k [Data set]. Hugging Face. https://huggingface.co/datasets/lmarena-ai/arena-human-preference-55k
LicenceApache 2.0
Source datahttps://huggingface.co/datasets/lmarena-ai/arena-human-preference-55k

Size and shape

Comparisons57,477
Agents64

Columns

agent_aagent_bscore_ascore_bwinner

Get the data

Download CSVno account neededBrowse on Redivisexplore and queryCroissant metadataHugging Face, Kaggle, OpenML

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_55k", source = "comp")
# Python
pip install irw

import irw
df = irw.fetch("lmarena_text_55k", 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_55k, clear source(comp)

How to cite

@dataset{lmarena2024arena,
  title={arena-human-preference-55k},
  author={LMArena},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/lmarena-ai/arena-human-preference-55k}
}

Version and provenance

IRW versionv712
Redivis datasetirw_competitions v7.1
Redivis dataset DOI10.57761/3r1w-t660
Manifest pin for this IRW versionv7.1
Metadata sourceirw_meta v68.0