|
|
import pandas as pd |
|
|
import streamlit as st |
|
|
|
|
|
st.set_page_config( |
|
|
page_title="JuStRank", |
|
|
page_icon="️🧑🏻⚖️", |
|
|
|
|
|
initial_sidebar_state="auto", |
|
|
menu_items=None, |
|
|
) |
|
|
|
|
|
st.title("🧑🏻⚖️ JuStRank: The Best Judges for Ranking Systems 🧑🏻⚖️") |
|
|
|
|
|
url = "https://arxiv.org/abs/2412.09569" |
|
|
st.subheader("Check out our [ACL paper](%s) for more details" % url) |
|
|
|
|
|
def prettify_judge_name(judge_name): |
|
|
pretty_judge = (judge_name[0].upper()+judge_name[1:]).replace("Gpt", "GPT") |
|
|
return pretty_judge |
|
|
|
|
|
|
|
|
def format_digits(flt, num_digits=3): |
|
|
format_str = "{:."+str(num_digits-1)+"f}" |
|
|
format_str_zeroes = "{:."+str(num_digits)+"f}" |
|
|
return format_str_zeroes.format(flt)[1:] if (0 < flt < 1) else format_str.format(flt) |
|
|
|
|
|
|
|
|
df = pd.read_csv("./best_judges_single_agg.csv")[["Judge Model", "Realization", "Ranking Agreement", "Decisiveness", "Bias"]] |
|
|
df["Judge Model"] = df["Judge Model"].apply(prettify_judge_name) |
|
|
|
|
|
styled_data = ( |
|
|
df.style.background_gradient(subset=["Ranking Agreement"]) |
|
|
.background_gradient( |
|
|
subset=["Ranking Agreement"], |
|
|
cmap="RdYlGn", |
|
|
vmin=0.5, |
|
|
vmax=0.9, |
|
|
) |
|
|
.format(subset=["Ranking Agreement", "Decisiveness", "Bias"], formatter=format_digits) |
|
|
.set_properties(**{"text-align": "center"}) |
|
|
) |
|
|
|
|
|
|
|
|
st.dataframe(styled_data, use_container_width=True, height=800, hide_index=True) |
|
|
|
|
|
st.text("\n\n") |
|
|
st.markdown( |
|
|
r""" |
|
|
This leaderboard measures the **system-level performance and behavior of LLM judges**, and was created as part of the **[JuStRank paper](https://www.arxiv.org/abs/2412.09569)** from ACL 2025. |
|
|
|
|
|
Judges are sorted according to *Ranking Agreement* with humans, i.e., comparing how the judges rank different systems (generative models) relative to how humans rank those systems on [LMSys Arena](https://lmarena.ai/leaderboard/text/hard-prompts-english). |
|
|
|
|
|
We also compare judges in terms of the *Decisiveness* and *Bias* reflected in their judgment behaviors (refer to the paper for details). |
|
|
|
|
|
In our research we tested 10 **LLM judges** and 8 **reward models**, and asked them to score the [responses](https://huggingface.co/datasets/lmarena-ai/arena-hard-auto/tree/main/data/arena-hard-v0.1/model_answer) of 63 systems to the 500 questions from Arena Hard v0.1. |
|
|
For each LLM judge we tried 4 different _realizations_, i.e., different prompt and scoring methods used with the LLM judge. |
|
|
|
|
|
In total, the judge ranking is derived from **[1.5 million raw judgment scores](https://huggingface.co/datasets/ibm-research/justrank_judge_scores)** (48 judge realizations X 63 target systems X 500 instances). |
|
|
|
|
|
If you find this useful, please cite our work 🤗 |
|
|
|
|
|
```bibtex |
|
|
@inproceedings{gera2025justrank, |
|
|
title={JuStRank: Benchmarking LLM Judges for System Ranking}, |
|
|
author={Gera, Ariel and Boni, Odellia and Perlitz, Yotam and Bar-Haim, Roy and Eden, Lilach and Yehudai, Asaf}, |
|
|
booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, |
|
|
month={july}, |
|
|
address={Vienna, Austria}, |
|
|
year={2025} |
|
|
url={www.arxiv.org/abs/2412.09569}, |
|
|
} |
|
|
``` |
|
|
""" |
|
|
) |
|
|
|
|
|
|
