File size: 15,118 Bytes
1ef58ee d05d9d8 1ef58ee d05d9d8 1ef58ee d05d9d8 1ef58ee d05d9d8 f9b31f1 d05d9d8 1ef58ee d05d9d8 1ef58ee d05d9d8 1ef58ee d05d9d8 1ef58ee 8b5abf6 1ef58ee |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 |
"""Script to produce radial plots."""
from functools import partial
import plotly.graph_objects as go
import json
import numpy as np
from collections import defaultdict
import pandas as pd
from pydantic import BaseModel
import gradio as gr
import requests
import random
class Task(BaseModel):
"""Class to hold task information."""
name: str
metric: str
def __hash__(self):
return hash(self.name)
class Language(BaseModel):
"""Class to hold language information."""
code: str
name: str
def __hash__(self):
return hash(self.code)
class Dataset(BaseModel):
"""Class to hold dataset information."""
name: str
language: Language
task: Task
def __hash__(self):
return hash(self.name)
TEXT_CLASSIFICATION = Task(name="text classification", metric="mcc")
INFORMATION_EXTRACTION = Task(name="information extraction", metric="micro_f1_no_misc")
GRAMMAR = Task(name="grammar", metric="mcc")
QUESTION_ANSWERING = Task(name="question answering", metric="em")
SUMMARISATION = Task(name="summarisation", metric="bertscore")
KNOWLEDGE = Task(name="knowledge", metric="mcc")
REASONING = Task(name="reasoning", metric="mcc")
ALL_TASKS = [obj for obj in globals().values() if isinstance(obj, Task)]
DANISH = Language(code="da", name="Danish")
NORWEGIAN = Language(code="no", name="Norwegian")
SWEDISH = Language(code="sv", name="Swedish")
ICELANDIC = Language(code="is", name="Icelandic")
FAROESE = Language(code="fo", name="Faroese")
GERMAN = Language(code="de", name="German")
DUTCH = Language(code="nl", name="Dutch")
ENGLISH = Language(code="en", name="English")
ALL_LANGUAGES = {
obj.name: obj for obj in globals().values() if isinstance(obj, Language)
}
DATASETS = [
Dataset(name="swerec", language=SWEDISH, task=TEXT_CLASSIFICATION),
Dataset(name="angry-tweets", language=DANISH, task=TEXT_CLASSIFICATION),
Dataset(name="norec", language=NORWEGIAN, task=TEXT_CLASSIFICATION),
Dataset(name="sb10k", language=GERMAN, task=TEXT_CLASSIFICATION),
Dataset(name="dutch-social", language=DUTCH, task=TEXT_CLASSIFICATION),
Dataset(name="sst5", language=ENGLISH, task=TEXT_CLASSIFICATION),
Dataset(name="suc3", language=SWEDISH, task=INFORMATION_EXTRACTION),
Dataset(name="dansk", language=DANISH, task=INFORMATION_EXTRACTION),
Dataset(name="norne-nb", language=NORWEGIAN, task=INFORMATION_EXTRACTION),
Dataset(name="norne-nn", language=NORWEGIAN, task=INFORMATION_EXTRACTION),
Dataset(name="mim-gold-ner", language=ICELANDIC, task=INFORMATION_EXTRACTION),
Dataset(name="fone", language=FAROESE, task=INFORMATION_EXTRACTION),
Dataset(name="germeval", language=GERMAN, task=INFORMATION_EXTRACTION),
Dataset(name="conll-nl", language=DUTCH, task=INFORMATION_EXTRACTION),
Dataset(name="conll-en", language=ENGLISH, task=INFORMATION_EXTRACTION),
Dataset(name="scala-sv", language=SWEDISH, task=GRAMMAR),
Dataset(name="scala-da", language=DANISH, task=GRAMMAR),
Dataset(name="scala-nb", language=NORWEGIAN, task=GRAMMAR),
Dataset(name="scala-nn", language=NORWEGIAN, task=GRAMMAR),
Dataset(name="scala-is", language=ICELANDIC, task=GRAMMAR),
Dataset(name="scala-fo", language=FAROESE, task=GRAMMAR),
Dataset(name="scala-de", language=GERMAN, task=GRAMMAR),
Dataset(name="scala-nl", language=DUTCH, task=GRAMMAR),
Dataset(name="scala-en", language=ENGLISH, task=GRAMMAR),
Dataset(name="scandiqa-da", language=DANISH, task=QUESTION_ANSWERING),
Dataset(name="norquad", language=NORWEGIAN, task=QUESTION_ANSWERING),
Dataset(name="scandiqa-sv", language=SWEDISH, task=QUESTION_ANSWERING),
Dataset(name="nqii", language=ICELANDIC, task=QUESTION_ANSWERING),
Dataset(name="germanquad", language=GERMAN, task=QUESTION_ANSWERING),
Dataset(name="squad", language=ENGLISH, task=QUESTION_ANSWERING),
Dataset(name="squad-nl", language=DUTCH, task=QUESTION_ANSWERING),
Dataset(name="nordjylland-news", language=DANISH, task=SUMMARISATION),
Dataset(name="mlsum", language=GERMAN, task=SUMMARISATION),
Dataset(name="rrn", language=ICELANDIC, task=SUMMARISATION),
Dataset(name="no-sammendrag", language=NORWEGIAN, task=SUMMARISATION),
Dataset(name="wiki-lingua-nl", language=DUTCH, task=SUMMARISATION),
Dataset(name="swedn", language=SWEDISH, task=SUMMARISATION),
Dataset(name="cnn-dailymail", language=ENGLISH, task=SUMMARISATION),
Dataset(name="mmlu-da", language=DANISH, task=KNOWLEDGE),
Dataset(name="mmlu-no", language=NORWEGIAN, task=KNOWLEDGE),
Dataset(name="mmlu-sv", language=SWEDISH, task=KNOWLEDGE),
Dataset(name="mmlu-is", language=ICELANDIC, task=KNOWLEDGE),
Dataset(name="mmlu-de", language=GERMAN, task=KNOWLEDGE),
Dataset(name="mmlu-nl", language=DUTCH, task=KNOWLEDGE),
Dataset(name="mmlu", language=ENGLISH, task=KNOWLEDGE),
Dataset(name="arc-da", language=DANISH, task=KNOWLEDGE),
Dataset(name="arc-no", language=NORWEGIAN, task=KNOWLEDGE),
Dataset(name="arc-sv", language=SWEDISH, task=KNOWLEDGE),
Dataset(name="arc-is", language=ICELANDIC, task=KNOWLEDGE),
Dataset(name="arc-de", language=GERMAN, task=KNOWLEDGE),
Dataset(name="arc-nl", language=DUTCH, task=KNOWLEDGE),
Dataset(name="arc", language=ENGLISH, task=KNOWLEDGE),
Dataset(name="hellaswag-da", language=DANISH, task=REASONING),
Dataset(name="hellaswag-no", language=NORWEGIAN, task=REASONING),
Dataset(name="hellaswag-sv", language=SWEDISH, task=REASONING),
Dataset(name="hellaswag-is", language=ICELANDIC, task=REASONING),
Dataset(name="hellaswag-de", language=GERMAN, task=REASONING),
Dataset(name="hellaswag-nl", language=DUTCH, task=REASONING),
Dataset(name="hellaswag", language=ENGLISH, task=REASONING),
]
def main() -> None:
"""Produce a radial plot."""
# Download all the newest records
response = requests.get("https://scandeval.com/scandeval_benchmark_results.jsonl")
response.raise_for_status()
records = [
json.loads(dct_str)
for dct_str in response.text.split("\n")
if dct_str.strip("\n")
]
# Build a dictionary of languages -> results-dataframes, whose indices are the
# models and columns are the tasks.
results_dfs = dict()
for language in {dataset.language for dataset in DATASETS}:
possible_dataset_names = {
dataset.name for dataset in DATASETS if dataset.language == language
}
data_dict = defaultdict(dict)
for record in records:
model_name = record["model"]
dataset_name = record["dataset"]
if dataset_name in possible_dataset_names:
dataset = next(
dataset for dataset in DATASETS if dataset.name == dataset_name
)
results_dict = record['results']['total']
score = results_dict.get(
f"test_{dataset.task.metric}", results_dict.get(dataset.task.metric)
)
if dataset.task in data_dict[model_name]:
data_dict[model_name][dataset.task].append(score)
else:
data_dict[model_name][dataset.task] = [score]
results_df = pd.DataFrame(data_dict).T.map(
lambda list_or_nan:
np.mean(list_or_nan) if list_or_nan == list_or_nan else list_or_nan
).dropna()
if any(task not in results_df.columns for task in ALL_TASKS):
results_dfs[language] = pd.DataFrame()
else:
results_dfs[language] = results_df
all_languages: list[str | int | float | tuple[str, str | int | float]] | None = [
language.name for language in ALL_LANGUAGES.values()
]
all_models: list[str | int | float | tuple[str, str | int | float]] | None = list({
model_id
for df in results_dfs.values()
for model_id in df.index
})
with gr.Blocks(theme=gr.themes.Monochrome()) as demo:
gr.Markdown("# Radial Plot Generator")
gr.Markdown(
"This demo allows you to generate a radial plot comparing the performance "
"of different language models on different tasks. It is based on the "
"generative results from the [ScandEval benchmark](https://scandeval.com)."
)
with gr.Row():
with gr.Column():
language_names_dropdown = gr.Dropdown(
choices=all_languages,
multiselect=True,
label="Languages",
value=["Danish"],
interactive=True,
)
model_ids_dropdown = gr.Dropdown(
choices=all_models,
multiselect=True,
label="Models",
value=["gpt-3.5-turbo-0613", "mistralai/Mistral-7B-v0.1"],
interactive=True,
)
use_win_ratio_checkbox = gr.Checkbox(
label="Compare models with win ratios (as opposed to raw scores)",
value=True,
interactive=True,
)
gr.Markdown(
"<center>Made with ❤️ by the <a href=\"https://alexandra.dk\">"
"Alexandra Institute</a>.</center>"
)
with gr.Column():
plot = gr.Plot(
value=produce_radial_plot(
model_ids_dropdown.value,
language_names=language_names_dropdown.value,
use_win_ratio=use_win_ratio_checkbox.value,
results_dfs=results_dfs,
),
)
language_names_dropdown.change(
fn=partial(update_model_ids_dropdown, results_dfs=results_dfs),
inputs=language_names_dropdown,
outputs=model_ids_dropdown,
)
# Update plot when anything changes
language_names_dropdown.change(
fn=partial(produce_radial_plot, results_dfs=results_dfs),
inputs=[
model_ids_dropdown, language_names_dropdown, use_win_ratio_checkbox
],
outputs=plot,
)
model_ids_dropdown.change(
fn=partial(produce_radial_plot, results_dfs=results_dfs),
inputs=[
model_ids_dropdown, language_names_dropdown, use_win_ratio_checkbox
],
outputs=plot,
)
use_win_ratio_checkbox.change(
fn=partial(produce_radial_plot, results_dfs=results_dfs),
inputs=[
model_ids_dropdown, language_names_dropdown, use_win_ratio_checkbox
],
outputs=plot,
)
demo.launch()
def update_model_ids_dropdown(
language_names: list[str], results_dfs: dict[Language, pd.DataFrame] | None
) -> dict:
"""When the language names are updated, update the model ids dropdown.
Args:
language_names:
The names of the languages to include in the plot.
results_dfs:
The results dataframes for each language.
Returns:
The Gradio update to the model ids dropdown.
"""
if results_dfs is None or len(language_names) == 0:
return gr.update(choices=[], value=[])
filtered_results_dfs = {
language: df
for language, df in results_dfs.items()
if language.name in language_names
}
unique_models = {
model_id
for df in filtered_results_dfs.values()
for model_id in df.index
}
filtered_models = [
model_id
for model_id in unique_models
if all(model_id in df.index for df in filtered_results_dfs.values())
]
if len(filtered_models) == 0:
return gr.update(choices=[], value=[])
return gr.update(choices=filtered_models, value=random.sample(filtered_models, k=1))
def produce_radial_plot(
model_ids: list[str],
language_names: list[str],
use_win_ratio: bool,
results_dfs: dict[Language, pd.DataFrame] | None
) -> go.Figure:
"""Produce a radial plot as a plotly figure.
Args:
model_ids:
The ids of the models to include in the plot.
language_names:
The names of the languages to include in the plot.
use_win_ratio:
Whether to use win ratios (as opposed to raw scores).
results_dfs:
The results dataframes for each language.
Returns:
A plotly figure.
"""
if results_dfs is None or len(language_names) == 0 or len(model_ids) == 0:
return go.Figure()
tasks = ALL_TASKS
languages = [ALL_LANGUAGES[language_name] for language_name in language_names]
results_dfs_filtered = {
language: df
for language, df in results_dfs.items()
if language.name in language_names
}
# Add all the evaluation results for each model
results: list[list[float]] = list()
for model_id in model_ids:
result_list = list()
for task in tasks:
win_ratios = list()
scores = list()
for language in languages:
if model_id not in results_dfs_filtered[language].index:
continue
score = results_dfs_filtered[language].loc[model_id][task]
win_ratio = np.mean([
score >= other_score
for other_score in results_dfs_filtered[language][task].dropna()
])
win_ratios.append(win_ratio)
scores.append(score)
if use_win_ratio:
result_list.append(np.mean(win_ratios))
else:
result_list.append(np.mean(scores))
results.append(result_list)
# Sort the results to avoid misleading radial plots
model_idx_with_highest_variance = np.argmax(
[np.std(result_list) for result_list in results]
)
sorted_idxs = np.argsort(results[model_idx_with_highest_variance])
results = [np.asarray(result_list)[sorted_idxs] for result_list in results]
tasks = np.asarray(tasks)[sorted_idxs]
# Add the results to a plotly figure
fig = go.Figure()
for model_id, result_list in zip(model_ids, results):
fig.add_trace(go.Scatterpolar(
r=result_list,
theta=[task.name for task in tasks],
fill='toself',
name=model_id,
))
languages_str = ""
if len(languages) > 1:
languages_str = ", ".join([language.name for language in languages[:-1]])
languages_str += " and "
languages_str += languages[-1].name
if use_win_ratio:
title = f'Win Ratio on on {languages_str} Language Tasks'
else:
title = f'LLM Score on on {languages_str} Language Tasks'
# Builds the radial plot from the results
fig.update_layout(
polar=dict(radialaxis=dict(visible=True)), showlegend=True, title=title
)
return fig
if __name__ == "__main__":
main()
|