Spaces:
Running
Running
File size: 11,321 Bytes
e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 e9a40a3 ad3a876 5f36137 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 |
import gradio as gr
from collections import defaultdict
import os
import base64
import torch
from datasets import (
Dataset,
load_dataset,
)
import random
import pandas as pd
from collections import defaultdict
def encode_image_to_base64(image_path):
"""Encode an image or GIF file to base64."""
with open(image_path, "rb") as file:
encoded_string = base64.b64encode(file.read()).decode()
return encoded_string
def create_html_media(media_path, is_gif=False):
"""Create HTML for displaying an image or GIF."""
media_base64 = encode_image_to_base64(media_path)
media_type = "gif" if is_gif else "jpeg"
html_string = f"""
<div style="display: flex; justify-content: center; align-items: center; width: 100%; text-align: center;">
<div style="max-width: 450px; margin: auto;">
<img src="data:image/{media_type};base64,{media_base64}"
style="max-width: 75%; height: auto; display: block; margin: 0 auto; margin-top: 50px;"
alt="Displayed Media">
</div>
</div>
"""
return html_string
MASKED_LM_MODELS = [
"BounharAbdelaziz/XLM-RoBERTa-Morocco",
"SI2M-Lab/DarijaBERT",
"BounharAbdelaziz/ModernBERT-Morocco",
"google-bert/bert-base-multilingual-cased",
"FacebookAI/xlm-roberta-large",
"aubmindlab/bert-base-arabertv02",
]
CAUSAL_LM_MODELS = [
"BounharAbdelaziz/Al-Atlas-LLM-0.5B",
"Qwen/Qwen2.5-0.5B",
"tiiuae/Falcon3-1B-Base",
"MBZUAI-Paris/Atlas-Chat-2B",
]
class LMBattleArena:
def __init__(self, dataset_path):
"""Initialize battle arena with dataset"""
self.df = pd.read_csv(dataset_path)
print(self.df.head())
self.current_index = 0
self.saving_freq = 10 # save the results in csv/push to hub every 10 evaluations
self.evaluation_results_masked = []
self.evaluation_results_causal = []
self.model_scores = defaultdict(lambda: {'wins': 0, 'total_comparisons': 0})
def get_next_battle_pair(self, is_causal):
"""Retrieve next pair of summaries for comparison"""
if self.current_index >= len(self.df):
return None
row = self.df.iloc[self.current_index]
if is_causal:
model_summary_cols = [
col
for col in CAUSAL_LM_MODELS
]
else:
model_summary_cols = [
col
for col in MASKED_LM_MODELS
]
selected_models = random.sample(model_summary_cols, 2)
battle_data = {
'prompt': row['masked_sentence'] if not is_causal else row['causal_sentence'],
'model_1': row[selected_models[0]],
'model_2': row[selected_models[1]],
'model1_name': selected_models[0],
'model2_name': selected_models[1]
}
self.current_index += 1
return battle_data
def record_evaluation(self, preferred_models, input_text, output1, output2, model1_name, model2_name, is_causal):
"""Record user's model preference and update scores"""
self.model_scores[model1_name]['total_comparisons'] += 1
self.model_scores[model2_name]['total_comparisons'] += 1
if preferred_models == "Both Good":
self.model_scores[model1_name]['wins'] += 1
self.model_scores[model2_name]['wins'] += 1
elif preferred_models == "Model A": # Maps to first model
self.model_scores[model1_name]['wins'] += 1
elif preferred_models == "Model B": # Maps to second model
self.model_scores[model2_name]['wins'] += 1
# "Both Bad" case - no wins recorded
evaluation = {
'input_text': input_text,
'output1': output1,
'output2': output2,
'model1_name': model1_name,
'model2_name': model2_name,
'preferred_models': preferred_models
}
if is_causal:
self.evaluation_results_causal.append(evaluation)
else:
self.evaluation_results_masked.append(evaluation)
return self.get_model_scores_df(is_causal)
def get_model_scores_df(self, is_causal):
"""Convert model scores to DataFrame"""
scores_data = []
for model, stats in self.model_scores.items():
if is_causal:
if model not in CAUSAL_LM_MODELS:
continue
else:
if model not in MASKED_LM_MODELS:
continue
win_rate = (stats['wins'] / stats['total_comparisons'] * 100) if stats['total_comparisons'] > 0 else 0
scores_data.append({
'Model': model,
'Wins': stats['wins'],
'Total Comparisons': stats['total_comparisons'],
'Win Rate (%)': round(win_rate, 2)
})
results_df = pd.DataFrame(scores_data).sort_values('Win Rate (%)', ascending=False)
# save the results in a huggingface dataset
if self.current_index % self.saving_freq == 0 and self.current_index > 0:
# results_dataset = Dataset.from_pandas(results_df)
# results_dataset.push_to_hub('atlasia/Res-Moroccan-Darija-LLM-Battle-Al-Atlas', private=True)
results_df.to_csv('human_eval_results.csv')
return results_df
def create_battle_arena(dataset_path, is_gif, is_causal):
arena = LMBattleArena(dataset_path)
def battle_round(is_causal):
battle_data = arena.get_next_battle_pair(is_causal)
if battle_data is None:
return "No more texts to evaluate!", "", "", "", "", gr.DataFrame(visible=False)
return (
battle_data['prompt'],
battle_data['model_1'],
battle_data['model_2'],
battle_data['model1_name'],
battle_data['model2_name'],
gr.DataFrame(visible=True)
)
def submit_preference(input_text, output_1, output_2, model1_name, model2_name, preferred_models, is_causal):
scores_df = arena.record_evaluation(
preferred_models, input_text, output_1, output_2, model1_name, model2_name, is_causal
)
next_battle = battle_round(is_causal)
return (*next_battle[:-1], scores_df)
with gr.Blocks(css="footer{display:none !important}") as demo:
base_path = os.path.dirname(__file__)
local_image_path = os.path.join(base_path, 'battle_leaderboard.gif')
gr.HTML(create_html_media(local_image_path, is_gif=is_gif))
with gr.Tabs():
with gr.Tab("Masked LM Battle Arena"):
gr.Markdown("# 🤖 Pretrained SmolLMs Battle Arena")
# Use gr.State to store the boolean value without displaying it
is_causal = gr.State(value=False)
input_text = gr.Textbox(
label="Input prompt",
interactive=False,
)
with gr.Row():
output_1 = gr.Textbox(
label="Model A",
interactive=False
)
model1_name = gr.State() # Hidden state for model1 name
with gr.Row():
output_2 = gr.Textbox(
label="Model B",
interactive=False
)
model2_name = gr.State() # Hidden state for model2 name
preferred_models = gr.Radio(
label="Which model is better?",
choices=["Model A", "Model B", "Both Good", "Both Bad"]
)
submit_btn = gr.Button("Vote", variant="primary")
scores_table = gr.DataFrame(
headers=['Model', 'Wins', 'Total Comparisons', 'Win Rate (%)'],
label="🏆 Leaderboard"
)
submit_btn.click(
submit_preference,
inputs=[input_text, output_1, output_2, model1_name, model2_name, preferred_models, is_causal],
outputs=[input_text, output_1, output_2, model1_name, model2_name, scores_table]
)
demo.load(
battle_round,
inputs=[is_causal],
outputs=[input_text, output_1, output_2, model1_name, model2_name, scores_table]
)
with gr.Tab("Causal LM Battle Arena"):
gr.Markdown("# 🤖 Pretrained SmolLMs Battle Arena")
# Use gr.State to store the boolean value without displaying it
is_causal = gr.State(value=True)
input_text = gr.Textbox(
label="Input prompt",
interactive=False,
)
with gr.Row():
output_1 = gr.Textbox(
label="Model A",
interactive=False
)
model1_name = gr.State() # Hidden state for model1 name
with gr.Row():
output_2 = gr.Textbox(
label="Model B",
interactive=False
)
model2_name = gr.State() # Hidden state for model2 name
preferred_models = gr.Radio(
label="Which model is better?",
choices=["Model A", "Model B", "Both Good", "Both Bad"]
)
submit_btn = gr.Button("Vote", variant="primary")
scores_table = gr.DataFrame(
headers=['Model', 'Wins', 'Total Comparisons', 'Win Rate (%)'],
label="🏆 Leaderboard"
)
submit_btn.click(
submit_preference,
inputs=[input_text, output_1, output_2, model1_name, model2_name, preferred_models, is_causal],
outputs=[input_text, output_1, output_2, model1_name, model2_name, scores_table]
)
demo.load(
battle_round,
inputs=[is_causal],
outputs=[input_text, output_1, output_2, model1_name, model2_name, scores_table]
)
return demo
if __name__ == "__main__":
# load the existing dataset that contains outputs of the LMs
human_eval_dataset = load_dataset("atlasia/LM-Moroccan-Darija-Bench", split='test').to_csv('human_eval_dataset.csv')
# precision
torch_dtype = torch.float16
# inference device
device = "cpu" #"cuda" if torch.cuda.is_available() else "cpu"
dataset_path = 'human_eval_dataset.csv'
is_gif = True
demo = create_battle_arena(dataset_path, is_gif, is_causal=False)
demo.launch(debug=True) |