S-Eval_v0.1.1 / app.py
Yuanxh's picture
Update leaderboard
36713b7
raw
history blame
4.85 kB
import gradio as gr
__all__ = ["block", "make_clickable_model", "make_clickable_user", "get_submissions"]
import numpy as np
import pandas as pd
from constants import *
from src.auto_leaderboard.model_metadata_type import ModelType
global data_component, filter_component, ref_dic
def upload_file(files):
file_paths = [file.name for file in files]
return file_paths
def read_xlsx_leaderboard():
df_dict = pd.read_excel(XLSX_DIR, sheet_name=None) # get all sheet
return df_dict
def get_specific_df(sheet_name):
df = read_xlsx_leaderboard()[sheet_name].sort_values(by="Overall", ascending=False)
return df
def get_link_df(sheet_name):
df = read_xlsx_leaderboard()[sheet_name]
return df
ref_df = get_link_df("main")
ref_dic = {}
for id, row in ref_df.iterrows():
ref_dic[
str(row["Model"])
] = f'<a href="{row["Link"]}" target="_blank" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{row["Model"]}</a>'
def wrap_model(func):
def wrapper(*args, **kwargs):
df = func(*args, **kwargs)
df["Model"] = df["Model"].apply(lambda x: ref_dic[x])
cols_to_round = df.select_dtypes(include=[np.number]).columns.difference(
["Model"]
)
df[cols_to_round] = df[cols_to_round].apply(lambda x: np.round(x, 2))
return df
return wrapper
@wrap_model
def get_base_zh_df():
return get_specific_df("base-zh")
@wrap_model
def get_base_en_df():
return get_specific_df("base-en")
@wrap_model
def get_attack_zh_df():
return get_specific_df("attack-zh")
@wrap_model
def get_attack_en_df():
return get_specific_df("attack-en")
def build_leaderboard(
TABLE_INTRODUCTION, TAX_COLUMNS, get_chinese_df, get_english_df
):
gr.Markdown(TABLE_INTRODUCTION, elem_classes="markdown-text")
data_spilt_radio = gr.Radio(
choices=["Chinses", "English"],
value="base",
label=SELECT_SET_INTRO,
)
# 创建数据帧组件
data_component = gr.components.Dataframe(
value=get_chinese_df,
headers=OVERALL_INFO + TAX_COLUMNS,
type="pandas",
datatype=["markdown"] + ["number"] + ["number"] * len(TAX_COLUMNS),
interactive=False,
visible=True,
wrap=True,
column_widths=[300] + [100] + [200] * len(TAX_COLUMNS),
)
def on_data_split_radio(seleted_split):
if "Chinese" in seleted_split:
updated_data = get_chinese_df()
if "English" in seleted_split:
updated_data = get_english_df()
current_columns = data_component.headers # 获取的当前的column
current_datatype = data_component.datatype # 获取当前的datatype
filter_component = gr.components.Dataframe(
value=updated_data,
headers=current_columns,
type="pandas",
datatype=current_datatype,
interactive=False,
visible=True,
wrap=True,
column_widths=[300] + [100] + [200] * (len(current_columns) - 2),
)
return filter_component
# 关联处理函数
data_spilt_radio.change(
fn=on_data_split_radio, inputs=data_spilt_radio, outputs=data_component
)
def build_demo():
block = gr.Blocks()
with block:
gr.Markdown(LEADERBOARD_INTRODUCTION)
with gr.Tabs(elem_classes="tab-buttons") as tabs:
# first
with gr.TabItem(
"Base Risk Prompt Set Results",
elem_id="evalcrafter-benchmark-tab-table",
id=0,
):
build_leaderboard(
TABLE_INTRODUCTION_1,
get_base_zh_df,
get_base_en_df
)
# second
with gr.TabItem(
"Base Risk Prompt Set Results",
elem_id="evalcrafter-benchmark-tab-table",
id=1,
):
build_leaderboard(
TABLE_INTRODUCTION_2,
get_attack_zh_df,
get_attack_en_df
)
# last table about
with gr.TabItem("📝 About", elem_id="evalcrafter-benchmark-tab-table", id=3):
gr.Markdown(LEADERBORAD_INFO, elem_classes="markdown-text")
with gr.Row():
with gr.Accordion("📙 Citation", open=True):
citation_button = gr.Textbox(
value=CITATION_BUTTON_TEXT,
label=CITATION_BUTTON_LABEL,
lines=10,
elem_id="citation-button",
show_label=True,
show_copy_button=True,
)
block.launch(share=True)
if __name__ == "__main__":
build_demo()