Spaces:
Sleeping
Sleeping
File size: 11,781 Bytes
24d828a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 |
import gradio as gr
import numpy as np
import pandas as pd
import argparse
def make_default_md():
leaderboard_md = f"""
# 🏆 LLms Benchmark
The main goal of this project is to utilize Large Language Models (LLMs) to extract specific information from PDF documents and organize it into a structured JSON format.
To achieve this objective, we are assessing various LLMs on two benchmarks:
1. [Benchmark1](https://huggingface.co/spaces/Nechba/LLms-Benchmark/blob/main/dataset.jsonl):
This benchmark consists of a dataset of 59 pages as context and corresponding JSON extracts from "Interchange and Service Fees Manual: Europe Region".
2. [Benchmark2](https://huggingface.co/datasets/Effyis/Table-Extraction):
This benchmark comprises a dataset of 16573 tables as context and corresponding JSON extracts.
"""
return leaderboard_md
def make_arena_leaderboard_md(total_models):
leaderboard_md = f"""
Total #models: **{total_models}**. Last updated: Avril 04, 2024.
"""
return leaderboard_md
def model_hyperlink(model_name, link):
return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
def load_leaderboard_table_csv(filename, add_hyperlink=True):
rows = []
with open(filename, 'r') as file:
lines = file.readlines()
heads = [v.strip() for v in lines[0].split(",")]
for line in lines[1:]:
row = [v.strip() for v in line.split(",")]
item = {}
for h, v in zip(heads, row):
item[h] = v
if add_hyperlink:
item["Model"] = model_hyperlink(item["Model"], item["Link"])
item["Notebook link"] = model_hyperlink("Notebook", item["Notebook link"])
rows.append(item)
return rows
def get_arena_table(model_table_df):
# change type Percentage of values column of df
model_table_df["Percentage of values"] = model_table_df["Percentage of values"].astype(float)
model_table_df["Percentage of keys"] = model_table_df["Percentage of keys"].astype(float)
model_table_df["Average time (s)"] = model_table_df["Average time (s)"].astype(float)
arena_df = model_table_df.sort_values(by=["Percentage of values"], ascending=False)
values = []
if not arena_df.empty: # Check if arena_df is not empty
for i in range(len(arena_df)):
row = []
model_name = arena_df["Model"].values[i] # Access model name directly without index 0
row.append(model_name)
row.append(arena_df.iloc[i]["Percentage of values"])
row.append(arena_df.iloc[i]["Percentage of keys"])
row.append(arena_df.iloc[i]["Average time (s)"])
row.append(arena_df.iloc[i]["Notebook link"])
row.append(arena_df.iloc[i]["License"])
# row.append(arena_df.iloc[i]["Link"])
values.append(row)
return values
def build_leaderboard_tab(leaderboard_table_file1,leaderboard_table_file2, show_plot=False):
default_md = make_default_md()
md_1 = gr.Markdown(default_md, elem_id="leaderboard_markdown")
if leaderboard_table_file1:
data1 = load_leaderboard_table_csv(leaderboard_table_file1)
model_table_df1 = pd.DataFrame(data1)
data2 = load_leaderboard_table_csv(leaderboard_table_file2)
model_table_df2 = pd.DataFrame(data2)
with gr.Tabs() as tabs:
with gr.Tab(" 🏅 Benchmark 1", id=0):
arena_table_vals = get_arena_table(model_table_df1)
md = make_arena_leaderboard_md(len(arena_table_vals))
gr.Markdown(md, elem_id="leaderboard_markdown")
gr.Dataframe(
headers=[
"Model",
"Percentage of values (%)",
"Percentage of keys (%)",
"Average time (s)",
"Code",
"License",
],
datatype=[
"markdown",
"number",
"number",
"number",
"markdown",
"str"
],
value=arena_table_vals,
elem_id="arena_leaderboard_dataframe",
height=700,
column_widths=[200, 150, 150, 130, 100, 140],
wrap=True,
)
# Displaying a note about the leaderboard analysis
gr.Markdown(
f"""Note: Upon reviewing the leaderboard, it's evident that two models, Gemini and OpenHermes, outperform the others. Our next step involves a detailed analysis and comparison of the results obtained by these two models.""",
elem_id="leaderboard_markdown"
)
# Displaying additional statistics for Gemini and OpenHermes
gr.Markdown(
f"""## More Statistics for Gemini and OpenHermes\n
Now we will focus on Gemini and OpenHermes, diving deeper into their performance for a comprehensive comparison.""",
elem_id=0
)
# Displaying the confusion matrices for Gemini and OpenHermes
with gr.Row():
with gr.Column():
gr.Markdown(
"#### Figure 1: Gemini Confusion Matrix"
)
plot_1 = gr.Image("./Benchmark1/gemini_cm.png", show_label=False)
# Detailed analysis of Gemini's performance
gr.Markdown(
"""### True Positives:
Our model correctly identified all 18 pages lacking the desired information (Payment product, FeeTier, and Rate).
### True Negatives:
The model successfully predicted desired information on 39 out of 41 pages with an accuracy ranging from 12% to 100%. (For more details about accuracy, check the Notebook [here](https://huggingface.co/spaces/Effyis/LLms-Benchmark/blob/main/Benchmark1/gemini.ipynb))
### False Negatives:
In 2 instances, the model incorrectly predicted that pages lacked the desired information when they actually contained it.
### False Positives:
The model incorrectly predicted that 0 pages contained the desired information when they were actually missing it."""
)
with gr.Column():
gr.Markdown(
"#### Figure 2: OpenHermes Confusion Matrix"
)
plot_2 = gr.Image("./Benchmark1/openhermes_cm.png", show_label=False)
# Detailed analysis of OpenHermes's performance
gr.Markdown(
"""### True Positives:
Our model correctly identified 12 out of 18 pages lacking the desired information (Payment product, FeeTier, and Rate).
### True Negatives:
The model successfully predicted desired information on 21 out of 41 pages with an accuracy ranging from 5% to 66%. (For more details about accuracy, check the Notebook [here](https://huggingface.co/spaces/Effyis/LLms-Benchmark/blob/main/Benchmark1/openhermes.ipynb))
### False Negatives:
In 20 instances, the model incorrectly predicted that pages lacked the desired information when they actually contained it.
### False Positives:
The model incorrectly predicted that 6 pages contained the desired information when they were actually missing it."""
)
# Conclusion based on the analysis
gr.Markdown(
"""## Conclusion\n
Upon analyzing the performance of Gemini and OpenHermes, it becomes evident that both models exhibit strengths and weaknesses. Gemini demonstrates higher accuracy in identifying pages lacking desired information and also performs better in predicting pages containing the desired information. On the other hand, while OpenHermes shows good results in identifying pages lacking desired information, it achieves only 50% accuracy in predicting pages containing the desired information. Further fine-tuning of both models could lead to enhanced overall performance."""
)
with gr.Tab("🏅 Benchmark 2", id=1):
arena_table_vals = get_arena_table(model_table_df2)
md = make_arena_leaderboard_md(len(arena_table_vals))
gr.Markdown(md, elem_id="leaderboard_markdown")
gr.Dataframe(
headers=[
"Model",
"Percentage of values (%)",
"Percentage of keys (%)",
"Average time (s)",
"Code",
"License",
],
datatype=[
"markdown",
"number",
"number",
"number",
"markdown",
"str"
],
value=arena_table_vals,
elem_id="arena_leaderboard_dataframe",
height=700,
column_widths=[200, 150, 150, 130, 100, 140],
wrap=True,
)
gr.Markdown(
f"""
Note: For this benchmark, only a sample of 100 points from the dataset is utilized. It's evident that the data context is straightforward, yet it includes Arabic names. This could explain the lower performance scores of the models, as they may lack robust capabilities in handling Arabic names.""",
elem_id="leaderboard_markdown"
)
else:
pass
return [md_1,plot_1, plot_2]
block_css = """
#notice_markdown {
font-size: 104%
}
#notice_markdown th {
display: none;
}
#notice_markdown td {
padding-top: 6px;
padding-bottom: 6px;
}
#leaderboard_markdown {
font-size: 104%
}
#leaderboard_markdown td {
padding-top: 6px;
padding-bottom: 6px;
}
#leaderboard_dataframe td {
line-height: 0.1em;
}
footer {
display:none !important
}
.sponsor-image-about img {
margin: 0 20px;
margin-top: 20px;
height: 40px;
max-height: 100%;
width: auto;
float: left;
}
"""
def build_demo(leaderboard_table_file1, leaderboard_table_file2):
text_size = gr.themes.sizes.text_lg
with gr.Blocks(
title="LLMS Benchmark",
theme=gr.themes.Base(text_size=text_size),
css=block_css,
) as demo:
leader_components = build_leaderboard_tab(
leaderboard_table_file1,leaderboard_table_file2, show_plot=True
)
return demo
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--share", action="store_true")
args = parser.parse_args()
leaderboard_table_file1 = "./Benchmark1/leaderboard.csv"
leaderboard_table_file2 = "./Benchmark2/leaderboard.csv"
demo = build_demo(leaderboard_table_file1,leaderboard_table_file2)
demo.launch(share=args.share) |