Corey
commited on
Commit
•
59c6dd2
1
Parent(s):
cb2c32e
Added clickable links (#1)
Browse files* Added clickable links
* Combined organization and model as a single column name
* removed redundant filtering
* removed model comparison feature because it did not seem useful as it was made
* Updated to streamlit 1.25.0 for clickable link support
* Removed truthfulQA data temporarily.
- app.py +47 -63
- generate_csv.ipynb +63 -0
- processed_data_2023-10-06.csv +0 -0
- requirements.txt +1 -1
- result_data_processor.py +58 -26
app.py
CHANGED
@@ -95,7 +95,7 @@ def create_line_chart(df, model_names, metrics):
|
|
95 |
fig.update_layout(showlegend=True)
|
96 |
return fig
|
97 |
|
98 |
-
def find_top_differences_table(df, target_model, closest_models, num_differences=10, exclude_columns=['Parameters'
|
99 |
# Calculate the absolute differences for each task between the target model and the closest models
|
100 |
new_df = df.drop(columns=exclude_columns)
|
101 |
differences = new_df.loc[closest_models].sub(new_df.loc[target_model]).abs()
|
@@ -124,35 +124,12 @@ st.markdown("""
|
|
124 |
""")
|
125 |
|
126 |
# Load the data into memory
|
127 |
-
data_path = "processed_data_2023-10-
|
128 |
data_df = load_csv_data(data_path)
|
129 |
-
|
|
|
130 |
data_df.set_index("Model Name", inplace=True)
|
131 |
|
132 |
-
filters = st.checkbox('Select Models and/or Evaluations')
|
133 |
-
|
134 |
-
# Initialize selected columns with "Parameters" and "MMLU_average" if filters are checked
|
135 |
-
selected_columns = ['Parameters', 'MMLU_average'] if filters else data_df.columns.tolist()
|
136 |
-
|
137 |
-
# Initialize selected models as empty if filters are checked
|
138 |
-
selected_models = [] if filters else data_df.index.tolist()
|
139 |
-
|
140 |
-
if filters:
|
141 |
-
# Create multi-select for columns with default selection
|
142 |
-
selected_columns = st.multiselect(
|
143 |
-
'Select Columns',
|
144 |
-
data_df.columns.tolist(),
|
145 |
-
default=selected_columns
|
146 |
-
)
|
147 |
-
|
148 |
-
# Create multi-select for models without default selection
|
149 |
-
selected_models = st.multiselect(
|
150 |
-
'Select Models',
|
151 |
-
data_df.index.tolist()
|
152 |
-
)
|
153 |
-
|
154 |
-
# Get the filtered data
|
155 |
-
# filtered_data = data_provider.get_data(selected_models)
|
156 |
filtered_data = data_df
|
157 |
|
158 |
# sort the table by the MMLU_average column
|
@@ -165,32 +142,33 @@ parameter_threshold = st.selectbox(
|
|
165 |
index=4, # Set the default selected option to 'No threshold'
|
166 |
format_func=lambda x: f"{x}" if isinstance(x, int) else x
|
167 |
)
|
168 |
-
|
169 |
-
# Filter the DataFrame based on the selected parameter threshold if not 'No threshold'
|
170 |
if isinstance(parameter_threshold, int):
|
171 |
filtered_data = filtered_data[filtered_data['Parameters'] <= parameter_threshold]
|
172 |
|
|
|
|
|
|
|
|
|
173 |
|
174 |
-
#
|
175 |
-
search_query = st.text_input("Filter by Model Name:", "")
|
176 |
-
|
177 |
-
# Filter the DataFrame based on the search query in the index (model name)
|
178 |
-
if search_query:
|
179 |
-
filtered_data = filtered_data[filtered_data.index.str.contains(search_query, case=False)]
|
180 |
-
|
181 |
-
|
182 |
-
# Search box for columns
|
183 |
column_search_query = st.text_input("Filter by Column/Task Name:", "").replace(" ", "").split(',')
|
184 |
-
|
185 |
-
# Get the columns that contain the search query
|
186 |
matching_columns = [col for col in filtered_data.columns if any(query.lower() in col.lower() for query in column_search_query)]
|
|
|
|
|
187 |
|
188 |
# Display the DataFrame with only the matching columns
|
189 |
st.markdown("## Sortable Results")
|
190 |
-
st.dataframe(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
# CSV download
|
193 |
-
|
194 |
filtered_data.index.name = "Model Name"
|
195 |
|
196 |
csv = filtered_data.to_csv(index=True)
|
@@ -209,6 +187,9 @@ def create_plot(df, x_values, y_values, models=None, title=None):
|
|
209 |
# remove rows with NaN values
|
210 |
df = df.dropna(subset=[x_values, y_values])
|
211 |
|
|
|
|
|
|
|
212 |
plot_data = pd.DataFrame({
|
213 |
'Model': df.index,
|
214 |
x_values: df[x_values],
|
@@ -279,8 +260,11 @@ st.markdown("***The dashed red line indicates random chance accuracy of 0.25 as
|
|
279 |
st.markdown("***")
|
280 |
st.write("As expected, there is a strong positive relationship between the number of parameters and average performance on the MMLU evaluation.")
|
281 |
|
282 |
-
|
283 |
-
|
|
|
|
|
|
|
284 |
|
285 |
if selected_x_column != selected_y_column: # Avoid creating a plot with the same column on both axes
|
286 |
fig = create_plot(filtered_data, selected_x_column, selected_y_column)
|
@@ -289,44 +273,44 @@ else:
|
|
289 |
st.write("Please select different columns for the x and y axes.")
|
290 |
|
291 |
|
|
|
292 |
|
293 |
|
294 |
-
# end of custom scatter plots
|
295 |
|
296 |
-
# Section to select a model and display radar and line charts
|
297 |
-
st.header("Compare a Selected Model to the 5 Models Closest in MMLU Average Performance")
|
298 |
-
st.write("""
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
|
305 |
-
default_model_name = "GPT-JT-6B-v0"
|
306 |
|
307 |
-
default_model_index = filtered_data.index.tolist().index(default_model_name) if default_model_name in filtered_data.index else 0
|
308 |
-
selected_model_name = st.selectbox("Select a Model:", filtered_data.index.tolist(), index=default_model_index)
|
309 |
|
310 |
-
# Get the closest 5 models with unique indices
|
311 |
-
closest_models_diffs = filtered_data['MMLU_average'].sub(filtered_data.loc[selected_model_name, 'MMLU_average']).abs()
|
312 |
-
closest_models = closest_models_diffs.nsmallest(5, keep='first').index.drop_duplicates().tolist()
|
313 |
|
314 |
|
315 |
# Find the top 10 tasks with the largest differences and convert to a DataFrame
|
316 |
-
top_differences_table, top_differences_tasks = find_top_differences_table(filtered_data, selected_model_name, closest_models)
|
317 |
|
318 |
# Display the DataFrame for the closest models and the top differences tasks
|
319 |
-
st.dataframe(filtered_data.loc[closest_models, top_differences_tasks])
|
320 |
|
321 |
# # Display the table in the Streamlit app
|
322 |
# st.markdown("## Top Differences")
|
323 |
# st.dataframe(top_differences_table)
|
324 |
|
325 |
# Create a radar chart for the tasks with the largest differences
|
326 |
-
fig_radar_top_differences = create_radar_chart_unfilled(filtered_data, closest_models, top_differences_tasks)
|
327 |
|
328 |
# Display the radar chart
|
329 |
-
st.plotly_chart(fig_radar_top_differences)
|
330 |
|
331 |
|
332 |
st.markdown("## Notable findings and plots")
|
|
|
95 |
fig.update_layout(showlegend=True)
|
96 |
return fig
|
97 |
|
98 |
+
def find_top_differences_table(df, target_model, closest_models, num_differences=10, exclude_columns=['Parameters']):
|
99 |
# Calculate the absolute differences for each task between the target model and the closest models
|
100 |
new_df = df.drop(columns=exclude_columns)
|
101 |
differences = new_df.loc[closest_models].sub(new_df.loc[target_model]).abs()
|
|
|
124 |
""")
|
125 |
|
126 |
# Load the data into memory
|
127 |
+
data_path = "processed_data_2023-10-08.csv"
|
128 |
data_df = load_csv_data(data_path)
|
129 |
+
# drop the column Unnamed: 0
|
130 |
+
data_df.rename(columns={'Unnamed: 0': "Model Name"}, inplace=True)
|
131 |
data_df.set_index("Model Name", inplace=True)
|
132 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
filtered_data = data_df
|
134 |
|
135 |
# sort the table by the MMLU_average column
|
|
|
142 |
index=4, # Set the default selected option to 'No threshold'
|
143 |
format_func=lambda x: f"{x}" if isinstance(x, int) else x
|
144 |
)
|
|
|
|
|
145 |
if isinstance(parameter_threshold, int):
|
146 |
filtered_data = filtered_data[filtered_data['Parameters'] <= parameter_threshold]
|
147 |
|
148 |
+
# model name filtering
|
149 |
+
search_queries = st.text_input("Filter by Model Name:", "").replace(" ", "").split(',')
|
150 |
+
if search_queries:
|
151 |
+
filtered_data = filtered_data[filtered_data.index.str.contains('|'.join(search_queries), case=False)]
|
152 |
|
153 |
+
# column name filtering
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
column_search_query = st.text_input("Filter by Column/Task Name:", "").replace(" ", "").split(',')
|
|
|
|
|
155 |
matching_columns = [col for col in filtered_data.columns if any(query.lower() in col.lower() for query in column_search_query)]
|
156 |
+
filtered_data = filtered_data[matching_columns]
|
157 |
+
|
158 |
|
159 |
# Display the DataFrame with only the matching columns
|
160 |
st.markdown("## Sortable Results")
|
161 |
+
st.dataframe(
|
162 |
+
filtered_data[matching_columns],
|
163 |
+
column_config={
|
164 |
+
"URL": st.column_config.LinkColumn( # Only current way to make url a clickable link with streamlit without removing the interactivity of the table
|
165 |
+
width="small"
|
166 |
+
)
|
167 |
+
},
|
168 |
+
hide_index=True,
|
169 |
+
)
|
170 |
|
171 |
# CSV download
|
|
|
172 |
filtered_data.index.name = "Model Name"
|
173 |
|
174 |
csv = filtered_data.to_csv(index=True)
|
|
|
187 |
# remove rows with NaN values
|
188 |
df = df.dropna(subset=[x_values, y_values])
|
189 |
|
190 |
+
#remove label rows URL, full_model_name
|
191 |
+
df = df.drop(columns=['URL', 'full_model_name'])
|
192 |
+
|
193 |
plot_data = pd.DataFrame({
|
194 |
'Model': df.index,
|
195 |
x_values: df[x_values],
|
|
|
260 |
st.markdown("***")
|
261 |
st.write("As expected, there is a strong positive relationship between the number of parameters and average performance on the MMLU evaluation.")
|
262 |
|
263 |
+
column_list_for_plotting = filtered_data.columns.tolist()
|
264 |
+
column_list_for_plotting.remove('URL')
|
265 |
+
column_list_for_plotting.remove('full_model_name')
|
266 |
+
selected_x_column = st.selectbox('Select x-axis', column_list_for_plotting, index=0)
|
267 |
+
selected_y_column = st.selectbox('Select y-axis', column_list_for_plotting, index=1)
|
268 |
|
269 |
if selected_x_column != selected_y_column: # Avoid creating a plot with the same column on both axes
|
270 |
fig = create_plot(filtered_data, selected_x_column, selected_y_column)
|
|
|
273 |
st.write("Please select different columns for the x and y axes.")
|
274 |
|
275 |
|
276 |
+
# end of custom scatter plots
|
277 |
|
278 |
|
|
|
279 |
|
280 |
+
# # Section to select a model and display radar and line charts
|
281 |
+
# st.header("Compare a Selected Model to the 5 Models Closest in MMLU Average Performance")
|
282 |
+
# st.write("""
|
283 |
+
# This comparison highlights the nuances in model performance across different tasks.
|
284 |
+
# While the overall MMLU average score provides a general understanding of a model's capabilities,
|
285 |
+
# examining the closest models reveals variations in performance on individual tasks.
|
286 |
+
# Such an analysis can uncover specific strengths and weaknesses and guide further exploration and improvement.
|
287 |
+
# """)
|
288 |
|
289 |
+
# default_model_name = "GPT-JT-6B-v0"
|
290 |
|
291 |
+
# default_model_index = filtered_data.index.tolist().index(default_model_name) if default_model_name in filtered_data.index else 0
|
292 |
+
# selected_model_name = st.selectbox("Select a Model:", filtered_data.index.tolist(), index=default_model_index)
|
293 |
|
294 |
+
# # Get the closest 5 models with unique indices
|
295 |
+
# closest_models_diffs = filtered_data['MMLU_average'].sub(filtered_data.loc[selected_model_name, 'MMLU_average']).abs()
|
296 |
+
# closest_models = closest_models_diffs.nsmallest(5, keep='first').index.drop_duplicates().tolist()
|
297 |
|
298 |
|
299 |
# Find the top 10 tasks with the largest differences and convert to a DataFrame
|
300 |
+
# top_differences_table, top_differences_tasks = find_top_differences_table(filtered_data, selected_model_name, closest_models)
|
301 |
|
302 |
# Display the DataFrame for the closest models and the top differences tasks
|
303 |
+
# st.dataframe(filtered_data.loc[closest_models, top_differences_tasks])
|
304 |
|
305 |
# # Display the table in the Streamlit app
|
306 |
# st.markdown("## Top Differences")
|
307 |
# st.dataframe(top_differences_table)
|
308 |
|
309 |
# Create a radar chart for the tasks with the largest differences
|
310 |
+
# fig_radar_top_differences = create_radar_chart_unfilled(filtered_data, closest_models, top_differences_tasks)
|
311 |
|
312 |
# Display the radar chart
|
313 |
+
# st.plotly_chart(fig_radar_top_differences)
|
314 |
|
315 |
|
316 |
st.markdown("## Notable findings and plots")
|
generate_csv.ipynb
ADDED
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": 1,
|
6 |
+
"metadata": {},
|
7 |
+
"outputs": [
|
8 |
+
{
|
9 |
+
"name": "stdout",
|
10 |
+
"output_type": "stream",
|
11 |
+
"text": [
|
12 |
+
"100\n",
|
13 |
+
"200\n",
|
14 |
+
"300\n",
|
15 |
+
"400\n",
|
16 |
+
"500\n",
|
17 |
+
"600\n",
|
18 |
+
"700\n",
|
19 |
+
"800\n",
|
20 |
+
"900\n",
|
21 |
+
"1000\n",
|
22 |
+
"1100\n",
|
23 |
+
"1200\n",
|
24 |
+
"1300\n",
|
25 |
+
"1400\n"
|
26 |
+
]
|
27 |
+
}
|
28 |
+
],
|
29 |
+
"source": [
|
30 |
+
"from result_data_processor import ResultDataProcessor\n",
|
31 |
+
"result = ResultDataProcessor()"
|
32 |
+
]
|
33 |
+
},
|
34 |
+
{
|
35 |
+
"cell_type": "code",
|
36 |
+
"execution_count": null,
|
37 |
+
"metadata": {},
|
38 |
+
"outputs": [],
|
39 |
+
"source": []
|
40 |
+
}
|
41 |
+
],
|
42 |
+
"metadata": {
|
43 |
+
"kernelspec": {
|
44 |
+
"display_name": "mmlu",
|
45 |
+
"language": "python",
|
46 |
+
"name": "python3"
|
47 |
+
},
|
48 |
+
"language_info": {
|
49 |
+
"codemirror_mode": {
|
50 |
+
"name": "ipython",
|
51 |
+
"version": 3
|
52 |
+
},
|
53 |
+
"file_extension": ".py",
|
54 |
+
"mimetype": "text/x-python",
|
55 |
+
"name": "python",
|
56 |
+
"nbconvert_exporter": "python",
|
57 |
+
"pygments_lexer": "ipython3",
|
58 |
+
"version": "3.10.12"
|
59 |
+
}
|
60 |
+
},
|
61 |
+
"nbformat": 4,
|
62 |
+
"nbformat_minor": 2
|
63 |
+
}
|
processed_data_2023-10-06.csv
CHANGED
The diff for this file is too large to render.
See raw diff
|
|
requirements.txt
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
# replicating the current hugging face streamlit and pandas versions
|
2 |
-
streamlit==1.
|
3 |
pandas==1.5.3
|
4 |
|
5 |
# Other
|
|
|
1 |
# replicating the current hugging face streamlit and pandas versions
|
2 |
+
streamlit==1.25.0
|
3 |
pandas==1.5.3
|
4 |
|
5 |
# Other
|
result_data_processor.py
CHANGED
@@ -96,32 +96,47 @@ class ResultDataProcessor:
|
|
96 |
|
97 |
|
98 |
def process_data(self):
|
99 |
-
|
|
|
100 |
dataframes = []
|
101 |
organization_names = []
|
102 |
for filename in self._find_files(self.directory, self.pattern):
|
103 |
-
try:
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
|
121 |
data = pd.concat(dataframes, axis=1).transpose()
|
122 |
|
123 |
# Add organization column
|
124 |
-
data['organization'] = organization_names
|
|
|
|
|
|
|
|
|
|
|
125 |
|
126 |
# Add Model Name and rearrange columns
|
127 |
data['Model Name'] = data.index
|
@@ -143,8 +158,7 @@ class ResultDataProcessor:
|
|
143 |
|
144 |
|
145 |
|
146 |
-
|
147 |
-
data = data.drop(columns=['all', 'truthfulqa:mc|0'])
|
148 |
|
149 |
# Add parameter count column using extract_parameters function
|
150 |
data['Parameters'] = data.index.to_series().apply(self._extract_parameters)
|
@@ -155,18 +169,36 @@ class ResultDataProcessor:
|
|
155 |
print(cols)
|
156 |
data = data[cols]
|
157 |
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
|
|
|
|
|
|
|
|
162 |
|
163 |
# remove extreme outliers from column harness|truthfulqa:mc1
|
164 |
-
data = self._remove_mc1_outliers(data)
|
165 |
|
166 |
data = self.manual_removal_of_models(data)
|
167 |
|
168 |
-
# save to csv with the current date as part of the filename
|
169 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
170 |
data.to_csv(f'processed_data_{pd.Timestamp.now().strftime("%Y-%m-%d")}.csv')
|
171 |
|
172 |
return data
|
|
|
96 |
|
97 |
|
98 |
def process_data(self):
|
99 |
+
full_model_name_count = 0
|
100 |
+
full_model_names = []
|
101 |
dataframes = []
|
102 |
organization_names = []
|
103 |
for filename in self._find_files(self.directory, self.pattern):
|
104 |
+
# try:
|
105 |
+
raw_data = self._read_and_transform_data(filename)
|
106 |
+
split_path = filename.split('/')
|
107 |
+
model_name = split_path[2]
|
108 |
+
organization_name = split_path[1]
|
109 |
+
full_model_name = f'{organization_name}/{model_name}'
|
110 |
+
full_model_name_count += 1
|
111 |
+
# print count every 100 models
|
112 |
+
if full_model_name_count % 100 == 0:
|
113 |
+
print(full_model_name_count)
|
114 |
+
|
115 |
+
cleaned_data = self._cleanup_dataframe(raw_data, model_name)
|
116 |
+
# mc1 = self._extract_mc1(raw_data, full_model_name)
|
117 |
+
# mc2 = self._extract_mc2(raw_data, full_model_name)
|
118 |
+
# cleaned_data = pd.concat([cleaned_data, mc1])
|
119 |
+
# cleaned_data = pd.concat([cleaned_data, mc2])
|
120 |
+
organization_names.append(organization_name)
|
121 |
+
full_model_names.append(full_model_name)
|
122 |
+
dataframes.append(cleaned_data)
|
123 |
+
# except Exception as e:
|
124 |
+
# # logging.error(f'Error processing {filename}')
|
125 |
+
# # logging.error(f'The error is: {e}')
|
126 |
+
# print(f'Error processing {filename}')
|
127 |
+
# print(f'The error is: {e}')
|
128 |
+
# continue
|
129 |
|
130 |
|
131 |
data = pd.concat(dataframes, axis=1).transpose()
|
132 |
|
133 |
# Add organization column
|
134 |
+
# data['organization'] = organization_names
|
135 |
+
print("full_model_names")
|
136 |
+
print(len(full_model_names))
|
137 |
+
print("organization_names")
|
138 |
+
print(len(organization_name))
|
139 |
+
data['full_model_name'] = full_model_names
|
140 |
|
141 |
# Add Model Name and rearrange columns
|
142 |
data['Model Name'] = data.index
|
|
|
158 |
|
159 |
|
160 |
|
161 |
+
|
|
|
162 |
|
163 |
# Add parameter count column using extract_parameters function
|
164 |
data['Parameters'] = data.index.to_series().apply(self._extract_parameters)
|
|
|
169 |
print(cols)
|
170 |
data = data[cols]
|
171 |
|
172 |
+
|
173 |
+
new_columns = ['full_model_name'] + [col for col in data.columns if col != 'full_model_name']
|
174 |
+
data = data.reindex(columns=new_columns)
|
175 |
+
|
176 |
+
# # Reorder columns to move 'organization' to the second position
|
177 |
+
# cols = data.columns.tolist()
|
178 |
+
# cols = cols[-1:] + cols[:-1]
|
179 |
+
# data = data[cols]
|
180 |
|
181 |
# remove extreme outliers from column harness|truthfulqa:mc1
|
182 |
+
# data = self._remove_mc1_outliers(data)
|
183 |
|
184 |
data = self.manual_removal_of_models(data)
|
185 |
|
|
|
186 |
|
187 |
+
# drop rows if MMLU_abstract_algebra is NaN
|
188 |
+
data = data.dropna(subset=['MMLU_abstract_algebra'])
|
189 |
+
|
190 |
+
# add a URL column that takes https://huggingface.co/ + full_model_name
|
191 |
+
data['URL'] = 'https://huggingface.co/' + data['full_model_name']
|
192 |
+
|
193 |
+
new_columns = ['URL'] + [col for col in data.columns if col != 'URL']
|
194 |
+
data = data.reindex(columns=new_columns)
|
195 |
+
|
196 |
+
# drop columns drop|3 gsm8k and winogrande
|
197 |
+
data = data.drop(columns=['drop|3', 'gsm8k', 'winogrande'])
|
198 |
+
# # Drop specific columns
|
199 |
+
data = data.drop(columns=['all', 'truthfulqa:mc|0'])
|
200 |
+
|
201 |
+
# save to csv with the current date as part of the filename
|
202 |
data.to_csv(f'processed_data_{pd.Timestamp.now().strftime("%Y-%m-%d")}.csv')
|
203 |
|
204 |
return data
|