XufengDuan commited on
Commit
56bf4e8
1 Parent(s): 06b8478

Update space

Browse files
README.md CHANGED
@@ -1,20 +1,32 @@
1
  ---
2
- title: HumanLikeness
3
  emoji: 🥇
4
- colorFrom: green
5
  colorTo: indigo
6
  sdk: gradio
 
7
  app_file: app.py
8
  pinned: true
9
- license: mit
 
 
 
 
10
  ---
11
 
12
- # Start the configuration
13
 
14
- Most of the variables to change for a default leaderboard are in `src/env.py` (replace the path for your leaderboard) and `src/about.py` (for tasks).
 
 
 
 
15
 
16
- Results files should have the following format and be stored as json files:
17
- ```json
 
 
 
 
18
  {
19
  "config": {
20
  "model_dtype": "torch.float16", # or torch.bfloat16 or 8bit or 4bit
@@ -32,13 +44,4 @@ Results files should have the following format and be stored as json files:
32
  }
33
  ```
34
 
35
- Request files are created automatically by this tool.
36
-
37
- If you encounter problem on the space, don't hesitate to restart it to remove the create eval-queue, eval-queue-bk, eval-results and eval-results-bk created folder.
38
-
39
- # Code logic for more complex edits
40
-
41
- You'll find
42
- - the main table' columns names and properties in `src/display/utils.py`
43
- - the logic to read all results and request files, then convert them in dataframe lines, in `src/leaderboard/read_evals.py`, and `src/populate.py`
44
- - teh logic to allow or filter submissions in `src/submission/submit.py` and `src/submission/check_validity.py`
 
1
  ---
2
+ title: Humanlike Evaluation Leaderboard
3
  emoji: 🥇
4
+ colorFrom: blue
5
  colorTo: indigo
6
  sdk: gradio
7
+ sdk_version: 4.37.1
8
  app_file: app.py
9
  pinned: true
10
+ license: apache-2.0
11
+ tags:
12
+ - leaderboard
13
+ models:
14
+ - google/gemma-2-9b
15
  ---
16
 
 
17
 
18
+ python>3.10
19
+ pip spacy
20
+ python -m spacy download en_core_web_sm
21
+ pip install google.generativeai
22
+ python -m spacy download en_core_web_trf
23
 
24
+ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
25
+
26
+ Most of the variables to change for a default leaderboard are in env (replace the path for your leaderboard) and src/display/about.
27
+
28
+ Results files should have the following format:
29
+ ```
30
  {
31
  "config": {
32
  "model_dtype": "torch.float16", # or torch.bfloat16 or 8bit or 4bit
 
44
  }
45
  ```
46
 
47
+ Request files are created automatically by this tool.
 
 
 
 
 
 
 
 
 
app.py CHANGED
@@ -1,110 +1,234 @@
1
  import gradio as gr
2
- from gradio_leaderboard import Leaderboard, ColumnFilter, SelectColumns
3
  import pandas as pd
4
  from apscheduler.schedulers.background import BackgroundScheduler
5
  from huggingface_hub import snapshot_download
6
 
7
- from src.about import (
8
- CITATION_BUTTON_LABEL,
9
- CITATION_BUTTON_TEXT,
10
- EVALUATION_QUEUE_TEXT,
11
- INTRODUCTION_TEXT,
12
- LLM_BENCHMARKS_TEXT,
13
- TITLE,
14
- )
15
  from src.display.css_html_js import custom_css
16
- from src.display.utils import (
17
- BENCHMARK_COLS,
18
- COLS,
19
- EVAL_COLS,
20
- EVAL_TYPES,
21
- AutoEvalColumn,
22
- ModelType,
23
- fields,
24
- WeightType,
25
- Precision
26
- )
27
- from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN
28
- from src.populate import get_evaluation_queue_df, get_leaderboard_df
29
- from src.submission.submit import add_new_eval
30
 
31
 
32
  def restart_space():
33
- API.restart_space(repo_id=REPO_ID)
34
 
35
- ### Space initialisation
36
  try:
37
- print(EVAL_REQUESTS_PATH)
38
  snapshot_download(
39
- repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
40
  )
41
  except Exception:
42
  restart_space()
43
  try:
44
- print(EVAL_RESULTS_PATH)
45
  snapshot_download(
46
- repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN
47
  )
48
  except Exception:
49
  restart_space()
50
 
51
-
52
- LEADERBOARD_DF = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS)
53
 
54
  (
55
  finished_eval_queue_df,
56
  running_eval_queue_df,
57
  pending_eval_queue_df,
58
- ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS)
59
-
60
- def init_leaderboard(dataframe):
61
- if dataframe is None or dataframe.empty:
62
- raise ValueError("Leaderboard DataFrame is empty or None.")
63
- return Leaderboard(
64
- value=dataframe,
65
- datatype=[c.type for c in fields(AutoEvalColumn)],
66
- select_columns=SelectColumns(
67
- default_selection=[c.name for c in fields(AutoEvalColumn) if c.displayed_by_default],
68
- cant_deselect=[c.name for c in fields(AutoEvalColumn) if c.never_hidden],
69
- label="Select Columns to Display:",
70
- ),
71
- search_columns=[AutoEvalColumn.model.name, AutoEvalColumn.license.name],
72
- hide_columns=[c.name for c in fields(AutoEvalColumn) if c.hidden],
73
- filter_columns=[
74
- ColumnFilter(AutoEvalColumn.model_type.name, type="checkboxgroup", label="Model types"),
75
- ColumnFilter(AutoEvalColumn.precision.name, type="checkboxgroup", label="Precision"),
76
- ColumnFilter(
77
- AutoEvalColumn.params.name,
78
- type="slider",
79
- min=0.01,
80
- max=150,
81
- label="Select the number of parameters (B)",
82
- ),
83
- ColumnFilter(
84
- AutoEvalColumn.still_on_hub.name, type="boolean", label="Deleted/incomplete", default=True
85
- ),
86
- ],
87
- bool_checkboxgroup_label="Hide models",
88
- interactive=False,
89
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90
 
91
 
92
  demo = gr.Blocks(css=custom_css)
93
  with demo:
94
- gr.HTML(TITLE)
95
- gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
96
 
97
  with gr.Tabs(elem_classes="tab-buttons") as tabs:
98
  with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
99
- leaderboard = init_leaderboard(LEADERBOARD_DF)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100
 
101
  with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
102
- gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
103
 
104
  with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3):
105
  with gr.Column():
106
  with gr.Row():
107
- gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
108
 
109
  with gr.Column():
110
  with gr.Accordion(
@@ -114,8 +238,8 @@ with demo:
114
  with gr.Row():
115
  finished_eval_table = gr.components.Dataframe(
116
  value=finished_eval_queue_df,
117
- headers=EVAL_COLS,
118
- datatype=EVAL_TYPES,
119
  row_count=5,
120
  )
121
  with gr.Accordion(
@@ -125,8 +249,8 @@ with demo:
125
  with gr.Row():
126
  running_eval_table = gr.components.Dataframe(
127
  value=running_eval_queue_df,
128
- headers=EVAL_COLS,
129
- datatype=EVAL_TYPES,
130
  row_count=5,
131
  )
132
 
@@ -137,8 +261,8 @@ with demo:
137
  with gr.Row():
138
  pending_eval_table = gr.components.Dataframe(
139
  value=pending_eval_queue_df,
140
- headers=EVAL_COLS,
141
- datatype=EVAL_TYPES,
142
  row_count=5,
143
  )
144
  with gr.Row():
@@ -149,7 +273,7 @@ with demo:
149
  model_name_textbox = gr.Textbox(label="Model name")
150
  revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
151
  model_type = gr.Dropdown(
152
- choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown],
153
  label="Model type",
154
  multiselect=False,
155
  value=None,
@@ -158,14 +282,14 @@ with demo:
158
 
159
  with gr.Column():
160
  precision = gr.Dropdown(
161
- choices=[i.value.name for i in Precision if i != Precision.Unknown],
162
  label="Precision",
163
  multiselect=False,
164
  value="float16",
165
  interactive=True,
166
  )
167
  weight_type = gr.Dropdown(
168
- choices=[i.value.name for i in WeightType],
169
  label="Weights type",
170
  multiselect=False,
171
  value="Original",
@@ -176,7 +300,7 @@ with demo:
176
  submit_button = gr.Button("Submit Eval")
177
  submission_result = gr.Markdown()
178
  submit_button.click(
179
- add_new_eval,
180
  [
181
  model_name_textbox,
182
  base_model_name_textbox,
@@ -191,8 +315,8 @@ with demo:
191
  with gr.Row():
192
  with gr.Accordion("📙 Citation", open=False):
193
  citation_button = gr.Textbox(
194
- value=CITATION_BUTTON_TEXT,
195
- label=CITATION_BUTTON_LABEL,
196
  lines=20,
197
  elem_id="citation-button",
198
  show_copy_button=True,
@@ -201,4 +325,4 @@ with demo:
201
  scheduler = BackgroundScheduler()
202
  scheduler.add_job(restart_space, "interval", seconds=1800)
203
  scheduler.start()
204
- demo.queue(default_concurrency_limit=40).launch()
 
1
  import gradio as gr
 
2
  import pandas as pd
3
  from apscheduler.schedulers.background import BackgroundScheduler
4
  from huggingface_hub import snapshot_download
5
 
6
+ import src.display.about as about
 
 
 
 
 
 
 
7
  from src.display.css_html_js import custom_css
8
+ import src.display.utils as utils
9
+ import src.envs as envs
10
+ import src.populate as populate
11
+ import src.submission.submit as submit
 
 
 
 
 
 
 
 
 
 
12
 
13
 
14
  def restart_space():
15
+ envs.API.restart_space(repo_id=envs.REPO_ID, token=envs.TOKEN)
16
 
 
17
  try:
18
+ print(envs.EVAL_REQUESTS_PATH)
19
  snapshot_download(
20
+ repo_id=envs.QUEUE_REPO, local_dir=envs.EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
21
  )
22
  except Exception:
23
  restart_space()
24
  try:
25
+ print(envs.EVAL_RESULTS_PATH)
26
  snapshot_download(
27
+ repo_id=envs.RESULTS_REPO, local_dir=envs.EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30
28
  )
29
  except Exception:
30
  restart_space()
31
 
32
+ raw_data, original_df = populate.get_leaderboard_df(envs.EVAL_RESULTS_PATH, envs.EVAL_REQUESTS_PATH, utils.COLS, utils.BENCHMARK_COLS)
33
+ leaderboard_df = original_df.copy()
34
 
35
  (
36
  finished_eval_queue_df,
37
  running_eval_queue_df,
38
  pending_eval_queue_df,
39
+ ) = populate.get_evaluation_queue_df(envs.EVAL_REQUESTS_PATH, utils.EVAL_COLS)
40
+
41
+
42
+ # Searching and filtering
43
+ def update_table(
44
+ hidden_df: pd.DataFrame,
45
+ columns: list,
46
+ type_query: list,
47
+ precision_query: str,
48
+ size_query: list,
49
+ show_deleted: bool,
50
+ query: str,
51
+ ):
52
+ filtered_df = filter_models(hidden_df, type_query, size_query, precision_query, show_deleted)
53
+ filtered_df = filter_queries(query, filtered_df)
54
+ df = select_columns(filtered_df, columns)
55
+ return df
56
+
57
+
58
+ def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame:
59
+ return df[(df[utils.AutoEvalColumn.dummy.name].str.contains(query, case=False))]
60
+
61
+
62
+ def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame:
63
+ always_here_cols = [
64
+ utils.AutoEvalColumn.model_type_symbol.name,
65
+ utils.AutoEvalColumn.model.name,
66
+ ]
67
+ # We use COLS to maintain sorting
68
+ filtered_df = df[
69
+ always_here_cols + [c for c in utils.COLS if c in df.columns and c in columns] + [utils.AutoEvalColumn.dummy.name]
70
+ ]
71
+ return filtered_df
72
+
73
+
74
+ def filter_queries(query: str, filtered_df: pd.DataFrame) -> pd.DataFrame:
75
+ final_df = []
76
+ if query != "":
77
+ queries = [q.strip() for q in query.split(";")]
78
+ for _q in queries:
79
+ _q = _q.strip()
80
+ if _q != "":
81
+ temp_filtered_df = search_table(filtered_df, _q)
82
+ if len(temp_filtered_df) > 0:
83
+ final_df.append(temp_filtered_df)
84
+ if len(final_df) > 0:
85
+ filtered_df = pd.concat(final_df)
86
+ filtered_df = filtered_df.drop_duplicates(
87
+ subset=[utils.AutoEvalColumn.model.name, utils.AutoEvalColumn.precision.name, utils.AutoEvalColumn.revision.name]
88
+ )
89
+
90
+ return filtered_df
91
+
92
+
93
+ def filter_models(
94
+ df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool
95
+ ) -> pd.DataFrame:
96
+ # Show all models
97
+ # if show_deleted:
98
+ # filtered_df = df
99
+ # else: # Show only still on the hub models
100
+ # filtered_df = df[df[utils.AutoEvalColumn.still_on_hub.name]]
101
+
102
+ filtered_df = df
103
+
104
+ type_emoji = [t[0] for t in type_query]
105
+ filtered_df = filtered_df.loc[df[utils.AutoEvalColumn.model_type_symbol.name].isin(type_emoji)]
106
+ filtered_df = filtered_df.loc[df[utils.AutoEvalColumn.precision.name].isin(precision_query + ["None"])]
107
+
108
+ numeric_interval = pd.IntervalIndex(sorted([utils.NUMERIC_INTERVALS[s] for s in size_query]))
109
+ params_column = pd.to_numeric(df[utils.AutoEvalColumn.params.name], errors="coerce")
110
+ mask = params_column.apply(lambda x: any(numeric_interval.contains(x)))
111
+ filtered_df = filtered_df.loc[mask]
112
+
113
+ return filtered_df
114
 
115
 
116
  demo = gr.Blocks(css=custom_css)
117
  with demo:
118
+ gr.HTML(about.TITLE)
119
+ gr.Markdown(about.INTRODUCTION_TEXT, elem_classes="markdown-text")
120
 
121
  with gr.Tabs(elem_classes="tab-buttons") as tabs:
122
  with gr.TabItem("🏅 LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0):
123
+ with gr.Row():
124
+ with gr.Column():
125
+ with gr.Row():
126
+ search_bar = gr.Textbox(
127
+ placeholder=" 🔍 Search for your model (separate multiple queries with `;`) and press ENTER...",
128
+ show_label=False,
129
+ elem_id="search-bar",
130
+ )
131
+ with gr.Row():
132
+ shown_columns = gr.CheckboxGroup(
133
+ choices=[
134
+ c.name
135
+ for c in utils.fields(utils.AutoEvalColumn)
136
+ if not c.hidden and not c.never_hidden and not c.dummy
137
+ ],
138
+ value=[
139
+ c.name
140
+ for c in utils.fields(utils.AutoEvalColumn)
141
+ if c.displayed_by_default and not c.hidden and not c.never_hidden
142
+ ],
143
+ label="Select columns to show",
144
+ elem_id="column-select",
145
+ interactive=True,
146
+ )
147
+ with gr.Row():
148
+ deleted_models_visibility = gr.Checkbox(
149
+ value=False, label="Show gated/private/deleted models", interactive=True
150
+ )
151
+ with gr.Column(min_width=320):
152
+ #with gr.Box(elem_id="box-filter"):
153
+ filter_columns_type = gr.CheckboxGroup(
154
+ label="Model types",
155
+ choices=[t.to_str() for t in utils.ModelType],
156
+ value=[t.to_str() for t in utils.ModelType],
157
+ interactive=True,
158
+ elem_id="filter-columns-type",
159
+ )
160
+ filter_columns_precision = gr.CheckboxGroup(
161
+ label="Precision",
162
+ choices=[i.value.name for i in utils.Precision],
163
+ value=[i.value.name for i in utils.Precision],
164
+ interactive=True,
165
+ elem_id="filter-columns-precision",
166
+ )
167
+ filter_columns_size = gr.CheckboxGroup(
168
+ label="Model sizes (in billions of parameters)",
169
+ choices=list(utils.NUMERIC_INTERVALS.keys()),
170
+ value=list(utils.NUMERIC_INTERVALS.keys()),
171
+ interactive=True,
172
+ elem_id="filter-columns-size",
173
+ )
174
+
175
+ leaderboard_table = gr.components.Dataframe(
176
+ value=leaderboard_df[
177
+ [c.name for c in utils.fields(utils.AutoEvalColumn) if c.never_hidden]
178
+ + shown_columns.value
179
+ + [utils.AutoEvalColumn.dummy.name]
180
+ ],
181
+ headers=[c.name for c in utils.fields(utils.AutoEvalColumn) if c.never_hidden] + shown_columns.value,
182
+ datatype=utils.TYPES,
183
+ elem_id="leaderboard-table",
184
+ interactive=False,
185
+ visible=True,
186
+ column_widths=["2%", "33%"]
187
+ )
188
+
189
+ # Dummy leaderboard for handling the case when the user uses backspace key
190
+ hidden_leaderboard_table_for_search = gr.components.Dataframe(
191
+ value=original_df[utils.COLS],
192
+ headers=utils.COLS,
193
+ datatype=utils.TYPES,
194
+ visible=False,
195
+ )
196
+ search_bar.submit(
197
+ update_table,
198
+ [
199
+ hidden_leaderboard_table_for_search,
200
+ shown_columns,
201
+ filter_columns_type,
202
+ filter_columns_precision,
203
+ filter_columns_size,
204
+ deleted_models_visibility,
205
+ search_bar,
206
+ ],
207
+ leaderboard_table,
208
+ )
209
+ for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility]:
210
+ selector.change(
211
+ update_table,
212
+ [
213
+ hidden_leaderboard_table_for_search,
214
+ shown_columns,
215
+ filter_columns_type,
216
+ filter_columns_precision,
217
+ filter_columns_size,
218
+ deleted_models_visibility,
219
+ search_bar,
220
+ ],
221
+ leaderboard_table,
222
+ queue=True,
223
+ )
224
 
225
  with gr.TabItem("📝 About", elem_id="llm-benchmark-tab-table", id=2):
226
+ gr.Markdown(about.LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
227
 
228
  with gr.TabItem("🚀 Submit here! ", elem_id="llm-benchmark-tab-table", id=3):
229
  with gr.Column():
230
  with gr.Row():
231
+ gr.Markdown(about.EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
232
 
233
  with gr.Column():
234
  with gr.Accordion(
 
238
  with gr.Row():
239
  finished_eval_table = gr.components.Dataframe(
240
  value=finished_eval_queue_df,
241
+ headers=utils.EVAL_COLS,
242
+ datatype=utils.EVAL_TYPES,
243
  row_count=5,
244
  )
245
  with gr.Accordion(
 
249
  with gr.Row():
250
  running_eval_table = gr.components.Dataframe(
251
  value=running_eval_queue_df,
252
+ headers=utils.EVAL_COLS,
253
+ datatype=utils.EVAL_TYPES,
254
  row_count=5,
255
  )
256
 
 
261
  with gr.Row():
262
  pending_eval_table = gr.components.Dataframe(
263
  value=pending_eval_queue_df,
264
+ headers=utils.EVAL_COLS,
265
+ datatype=utils.EVAL_TYPES,
266
  row_count=5,
267
  )
268
  with gr.Row():
 
273
  model_name_textbox = gr.Textbox(label="Model name")
274
  revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main")
275
  model_type = gr.Dropdown(
276
+ choices=[t.to_str(" : ") for t in utils.ModelType if t != utils.ModelType.Unknown],
277
  label="Model type",
278
  multiselect=False,
279
  value=None,
 
282
 
283
  with gr.Column():
284
  precision = gr.Dropdown(
285
+ choices=[i.value.name for i in utils.Precision if i != utils.Precision.Unknown],
286
  label="Precision",
287
  multiselect=False,
288
  value="float16",
289
  interactive=True,
290
  )
291
  weight_type = gr.Dropdown(
292
+ choices=[i.value.name for i in utils.WeightType],
293
  label="Weights type",
294
  multiselect=False,
295
  value="Original",
 
300
  submit_button = gr.Button("Submit Eval")
301
  submission_result = gr.Markdown()
302
  submit_button.click(
303
+ submit.add_new_eval,
304
  [
305
  model_name_textbox,
306
  base_model_name_textbox,
 
315
  with gr.Row():
316
  with gr.Accordion("📙 Citation", open=False):
317
  citation_button = gr.Textbox(
318
+ value=about.CITATION_BUTTON_TEXT,
319
+ label=about.CITATION_BUTTON_LABEL,
320
  lines=20,
321
  elem_id="citation-button",
322
  show_copy_button=True,
 
325
  scheduler = BackgroundScheduler()
326
  scheduler.add_job(restart_space, "interval", seconds=1800)
327
  scheduler.start()
328
+ demo.queue(default_concurrency_limit=40).launch()
requirements.txt CHANGED
@@ -1,16 +1,17 @@
1
- APScheduler
2
- black
3
- datasets
4
- gradio
5
- gradio[oauth]
6
- gradio_leaderboard==0.0.9
7
- gradio_client
8
  huggingface-hub>=0.18.0
9
- matplotlib
10
- numpy
11
- pandas
12
- python-dateutil
13
- tqdm
14
- transformers
 
 
15
  tokenizers>=0.15.0
16
- sentencepiece
 
1
+ APScheduler==3.10.1
2
+ black==23.11.0
3
+ click==8.1.3
4
+ datasets==2.14.5
5
+ gradio==4.4.0
6
+ gradio_client==0.7.0
 
7
  huggingface-hub>=0.18.0
8
+ litellm==1.15.1
9
+ matplotlib==3.7.1
10
+ numpy==1.24.2
11
+ pandas==2.0.0
12
+ python-dateutil==2.8.2
13
+ requests==2.28.2
14
+ tqdm==4.65.0
15
+ transformers==4.35.2
16
  tokenizers>=0.15.0
17
+ sentence-transformers==2.2.2
src/display/css_html_js.py CHANGED
@@ -33,11 +33,17 @@ custom_css = """
33
  background: none;
34
  border: none;
35
  }
36
-
37
  #search-bar {
38
  padding: 0px;
39
  }
40
 
 
 
 
 
 
 
41
  /* Limit the width of the first AutoEvalColumn so that names don't expand too much */
42
  table td:first-child,
43
  table th:first-child {
 
33
  background: none;
34
  border: none;
35
  }
36
+
37
  #search-bar {
38
  padding: 0px;
39
  }
40
 
41
+ /* Hides the final AutoEvalColumn */
42
+ #llm-benchmark-tab-table table td:last-child,
43
+ #llm-benchmark-tab-table table th:last-child {
44
+ display: none;
45
+ }
46
+
47
  /* Limit the width of the first AutoEvalColumn so that names don't expand too much */
48
  table td:first-child,
49
  table th:first-child {
src/display/formatting.py CHANGED
@@ -1,3 +1,12 @@
 
 
 
 
 
 
 
 
 
1
  def model_hyperlink(link, model_name):
2
  return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
3
 
 
1
+ import os
2
+ from datetime import datetime, timezone
3
+
4
+ from huggingface_hub import HfApi
5
+ from huggingface_hub.hf_api import ModelInfo
6
+
7
+
8
+ API = HfApi()
9
+
10
  def model_hyperlink(link, model_name):
11
  return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
12
 
src/display/utils.py CHANGED
@@ -3,7 +3,7 @@ from enum import Enum
3
 
4
  import pandas as pd
5
 
6
- from src.about import Tasks
7
 
8
  def fields(raw_class):
9
  return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
@@ -19,16 +19,18 @@ class ColumnContent:
19
  displayed_by_default: bool
20
  hidden: bool = False
21
  never_hidden: bool = False
 
22
 
23
  ## Leaderboard columns
24
  auto_eval_column_dict = []
25
  # Init
26
- auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
27
- auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
28
- #Scores
29
- auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
30
  for task in Tasks:
31
  auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
 
32
  # Model information
33
  auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
34
  auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
@@ -39,6 +41,8 @@ auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B
39
  auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
40
  auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)])
41
  auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
 
 
42
 
43
  # We use make dataclass to dynamically fill the scores from Tasks
44
  AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
@@ -91,6 +95,9 @@ class WeightType(Enum):
91
  class Precision(Enum):
92
  float16 = ModelDetails("float16")
93
  bfloat16 = ModelDetails("bfloat16")
 
 
 
94
  Unknown = ModelDetails("?")
95
 
96
  def from_str(precision):
@@ -98,13 +105,32 @@ class Precision(Enum):
98
  return Precision.float16
99
  if precision in ["torch.bfloat16", "bfloat16"]:
100
  return Precision.bfloat16
 
 
 
 
 
 
101
  return Precision.Unknown
102
 
103
  # Column selection
104
  COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
 
 
 
105
 
106
  EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
107
  EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
108
 
109
  BENCHMARK_COLS = [t.value.col_name for t in Tasks]
110
 
 
 
 
 
 
 
 
 
 
 
 
3
 
4
  import pandas as pd
5
 
6
+ from src.display.about import Tasks
7
 
8
  def fields(raw_class):
9
  return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]
 
19
  displayed_by_default: bool
20
  hidden: bool = False
21
  never_hidden: bool = False
22
+ dummy: bool = False
23
 
24
  ## Leaderboard columns
25
  auto_eval_column_dict = []
26
  # Init
27
+ auto_eval_column_dict.append(["model_type_symbol", ColumnContent,
28
+ ColumnContent("T", "str", True, never_hidden=True)])
29
+ auto_eval_column_dict.append(["model", ColumnContent,
30
+ ColumnContent("Model", "markdown", True, never_hidden=True)])
31
  for task in Tasks:
32
  auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])
33
+
34
  # Model information
35
  auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
36
  auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
 
41
  auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❤️", "number", False)])
42
  auto_eval_column_dict.append(["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False)])
43
  auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
44
+ # Dummy column for the search bar (hidden by the custom CSS)
45
+ auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "str", False, dummy=True)])
46
 
47
  # We use make dataclass to dynamically fill the scores from Tasks
48
  AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)
 
95
  class Precision(Enum):
96
  float16 = ModelDetails("float16")
97
  bfloat16 = ModelDetails("bfloat16")
98
+ qt_8bit = ModelDetails("8bit")
99
+ qt_4bit = ModelDetails("4bit")
100
+ qt_GPTQ = ModelDetails("GPTQ")
101
  Unknown = ModelDetails("?")
102
 
103
  def from_str(precision):
 
105
  return Precision.float16
106
  if precision in ["torch.bfloat16", "bfloat16"]:
107
  return Precision.bfloat16
108
+ if precision in ["8bit"]:
109
+ return Precision.qt_8bit
110
+ if precision in ["4bit"]:
111
+ return Precision.qt_4bit
112
+ if precision in ["GPTQ", "None"]:
113
+ return Precision.qt_GPTQ
114
  return Precision.Unknown
115
 
116
  # Column selection
117
  COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
118
+ TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
119
+ COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
120
+ TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
121
 
122
  EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
123
  EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]
124
 
125
  BENCHMARK_COLS = [t.value.col_name for t in Tasks]
126
 
127
+ NUMERIC_INTERVALS = {
128
+ "?": pd.Interval(-1, 0, closed="right"),
129
+ "~1.5": pd.Interval(0, 2, closed="right"),
130
+ "~3": pd.Interval(2, 4, closed="right"),
131
+ "~7": pd.Interval(4, 9, closed="right"),
132
+ "~13": pd.Interval(9, 20, closed="right"),
133
+ "~35": pd.Interval(20, 45, closed="right"),
134
+ "~60": pd.Interval(45, 70, closed="right"),
135
+ "70+": pd.Interval(70, 10000, closed="right"),
136
+ }
src/envs.py CHANGED
@@ -1,19 +1,25 @@
1
  import os
2
-
3
  from huggingface_hub import HfApi
4
 
5
- # Info to change for your repository
6
- # ----------------------------------
7
- TOKEN = os.environ.get("HF_TOKEN") # A read/write token for your org
8
 
9
- OWNER = "demo-leaderboard-backend" # Change to your org - don't forget to create a results and request dataset, with the correct format!
 
 
 
 
 
 
 
 
 
10
  # ----------------------------------
11
 
12
- REPO_ID = f"{OWNER}/leaderboard"
13
  QUEUE_REPO = f"{OWNER}/requests"
14
  RESULTS_REPO = f"{OWNER}/results"
15
 
16
- # If you setup a cache later, just change HF_HOME
17
  CACHE_PATH=os.getenv("HF_HOME", ".")
18
 
19
  # Local caches
@@ -21,5 +27,20 @@ EVAL_REQUESTS_PATH = os.path.join(CACHE_PATH, "eval-queue")
21
  EVAL_RESULTS_PATH = os.path.join(CACHE_PATH, "eval-results")
22
  EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")
23
  EVAL_RESULTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-results-bk")
24
-
 
 
25
  API = HfApi(token=TOKEN)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  import os
2
+ import torch
3
  from huggingface_hub import HfApi
4
 
 
 
 
5
 
6
+ # replace this with our token
7
+ TOKEN = os.environ.get("HF_TOKEN", None)
8
+ # print(TOKEN)
9
+ # OWNER = "vectara"
10
+ # REPO_ID = f"{OWNER}/Humanlike"
11
+ # QUEUE_REPO = f"{OWNER}/requests"
12
+ # RESULTS_REPO = f"{OWNER}/results"
13
+
14
+
15
+ OWNER = "Simondon" # Change to your org - don't forget to create a results and request dataset, with the correct format!
16
  # ----------------------------------
17
 
18
+ REPO_ID = f"{OWNER}/Humanlike"
19
  QUEUE_REPO = f"{OWNER}/requests"
20
  RESULTS_REPO = f"{OWNER}/results"
21
 
22
+ # print(RESULTS_REPO)
23
  CACHE_PATH=os.getenv("HF_HOME", ".")
24
 
25
  # Local caches
 
27
  EVAL_RESULTS_PATH = os.path.join(CACHE_PATH, "eval-results")
28
  EVAL_REQUESTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-queue-bk")
29
  EVAL_RESULTS_PATH_BACKEND = os.path.join(CACHE_PATH, "eval-results-bk")
30
+ # print(EVAL_RESULTS_PATH)
31
+ # exit()
32
+ DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') #"cpu"
33
  API = HfApi(token=TOKEN)
34
+
35
+ DATASET_PATH = "./src/datasets/Material_Llama2_0603.xlsx" #experiment data
36
+ PROMPT_PATH = "./src/datasets/prompt.xlsx" #prompt for each experiment
37
+ HEM_PATH = 'vectara/hallucination_evaluation_model'
38
+ HUMAN_DATA = "./src/datasets/human_data.csv" #experiment data
39
+ ITEM_4_DATA = "./src/datasets/associataion_dataset.csv" #database
40
+ ITEM_5_DATA = "./src/datasets/Items_5.csv" #experiment 5 need verb words
41
+
42
+ # SYSTEM_PROMPT = "You are a chat bot answering questions using data. You must stick to the answers provided solely by the text in the passage provided."
43
+ SYSTEM_PROMPT = "You are a participant of a psycholinguistic experiment. You will do a task on English language use."
44
+ '''prompt'''
45
+ # USER_PROMPT = "You are asked the question 'Provide a concise summary of the following passage, covering the core pieces of information described': "
46
+ USER_PROMPT = ""
src/leaderboard/read_evals.py CHANGED
@@ -1,35 +1,32 @@
1
  import glob
2
  import json
3
- import math
4
  import os
5
  from dataclasses import dataclass
6
 
7
- import dateutil
8
  import numpy as np
 
9
 
10
- from src.display.formatting import make_clickable_model
11
- from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
12
- from src.submission.check_validity import is_model_on_hub
13
 
14
 
15
  @dataclass
16
  class EvalResult:
17
- """Represents one full evaluation. Built from a combination of the result and request file for a given run.
18
- """
19
- eval_name: str # org_model_precision (uid)
20
- full_model: str # org/model (path on hub)
21
- org: str
22
  model: str
23
- revision: str # commit hash, "" if main
24
  results: dict
25
- precision: Precision = Precision.Unknown
26
- model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
27
- weight_type: WeightType = WeightType.Original # Original or Adapter
28
- architecture: str = "Unknown"
29
  license: str = "?"
30
  likes: int = 0
31
  num_params: int = 0
32
- date: str = "" # submission date of request file
33
  still_on_hub: bool = False
34
 
35
  @classmethod
@@ -41,43 +38,35 @@ class EvalResult:
41
  config = data.get("config")
42
 
43
  # Precision
44
- precision = Precision.from_str(config.get("model_dtype"))
45
 
46
  # Get model and org
47
- org_and_model = config.get("model_name", config.get("model_args", None))
48
- org_and_model = org_and_model.split("/", 1)
49
 
50
- if len(org_and_model) == 1:
51
- org = None
52
- model = org_and_model[0]
53
- result_key = f"{model}_{precision.value.name}"
54
- else:
55
- org = org_and_model[0]
56
- model = org_and_model[1]
57
  result_key = f"{org}_{model}_{precision.value.name}"
58
- full_model = "/".join(org_and_model)
 
59
 
60
- still_on_hub, _, model_config = is_model_on_hub(
61
- full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
62
- )
63
- architecture = "?"
64
- if model_config is not None:
65
- architectures = getattr(model_config, "architectures", None)
66
- if architectures:
67
- architecture = ";".join(architectures)
68
 
69
  # Extract results available in this file (some results are split in several files)
70
  results = {}
71
- for task in Tasks:
72
  task = task.value
73
 
74
  # We average all scores of a given metric (not all metrics are present in all files)
75
  accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
76
- if accs.size == 0 or any([acc is None for acc in accs]):
77
- continue
78
 
79
- mean_acc = np.mean(accs) * 100.0
80
- results[task.benchmark] = mean_acc
81
 
82
  return self(
83
  eval_name=result_key,
@@ -85,7 +74,7 @@ class EvalResult:
85
  org=org,
86
  model=model,
87
  results=results,
88
- precision=precision,
89
  revision= config.get("model_sha", ""),
90
  still_on_hub=still_on_hub,
91
  architecture=architecture
@@ -93,40 +82,43 @@ class EvalResult:
93
 
94
  def update_with_request_file(self, requests_path):
95
  """Finds the relevant request file for the current model and updates info with it"""
96
- request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)
 
97
 
98
  try:
99
  with open(request_file, "r") as f:
100
  request = json.load(f)
101
- self.model_type = ModelType.from_str(request.get("model_type", ""))
102
- self.weight_type = WeightType[request.get("weight_type", "Original")]
103
  self.license = request.get("license", "?")
104
  self.likes = request.get("likes", 0)
105
  self.num_params = request.get("params", 0)
106
  self.date = request.get("submitted_time", "")
107
- except Exception:
108
- print(f"Could not find request file for {self.org}/{self.model} with precision {self.precision.value.name}")
 
 
109
 
110
  def to_dict(self):
111
  """Converts the Eval Result to a dict compatible with our dataframe display"""
112
- average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
113
  data_dict = {
114
  "eval_name": self.eval_name, # not a column, just a save name,
115
- AutoEvalColumn.precision.name: self.precision.value.name,
116
- AutoEvalColumn.model_type.name: self.model_type.value.name,
117
- AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
118
- AutoEvalColumn.weight_type.name: self.weight_type.value.name,
119
- AutoEvalColumn.architecture.name: self.architecture,
120
- AutoEvalColumn.model.name: make_clickable_model(self.full_model),
121
- AutoEvalColumn.revision.name: self.revision,
122
- AutoEvalColumn.average.name: average,
123
- AutoEvalColumn.license.name: self.license,
124
- AutoEvalColumn.likes.name: self.likes,
125
- AutoEvalColumn.params.name: self.num_params,
126
- AutoEvalColumn.still_on_hub.name: self.still_on_hub,
127
  }
128
 
129
- for task in Tasks:
130
  data_dict[task.value.col_name] = self.results[task.value.benchmark]
131
 
132
  return data_dict
@@ -157,21 +149,26 @@ def get_request_file_for_model(requests_path, model_name, precision):
157
  def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
158
  """From the path of the results folder root, extract all needed info for results"""
159
  model_result_filepaths = []
160
-
161
  for root, _, files in os.walk(results_path):
162
  # We should only have json files in model results
163
- if len(files) == 0 or any([not f.endswith(".json") for f in files]):
164
- continue
165
 
166
- # Sort the files by date
167
- try:
168
- files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
169
- except dateutil.parser._parser.ParserError:
170
- files = [files[-1]]
171
-
172
- for file in files:
173
- model_result_filepaths.append(os.path.join(root, file))
174
 
 
 
 
 
 
 
 
 
 
175
  eval_results = {}
176
  for model_result_filepath in model_result_filepaths:
177
  # Creation of result
@@ -181,7 +178,8 @@ def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResu
181
  # Store results of same eval together
182
  eval_name = eval_result.eval_name
183
  if eval_name in eval_results.keys():
184
- eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
 
185
  else:
186
  eval_results[eval_name] = eval_result
187
 
 
1
  import glob
2
  import json
 
3
  import os
4
  from dataclasses import dataclass
5
 
 
6
  import numpy as np
7
+ import dateutil
8
 
9
+ import src.display.formatting as formatting
10
+ import src.display.utils as utils
11
+ import src.submission.check_validity as check_validity
12
 
13
 
14
  @dataclass
15
  class EvalResult:
16
+ eval_name: str # org_model_precision (uid)
17
+ full_model: str # org/model (path on hub)
18
+ org: str
 
 
19
  model: str
20
+ revision: str # commit hash, "" if main
21
  results: dict
22
+ precision: utils.Precision = utils.Precision.Unknown
23
+ model_type: utils.ModelType = utils.ModelType.Unknown # Pretrained, fine tuned, ...
24
+ weight_type: utils.WeightType = utils.WeightType.Original # Original or Adapter
25
+ architecture: str = "Unknown"
26
  license: str = "?"
27
  likes: int = 0
28
  num_params: int = 0
29
+ date: str = "" # submission date of request file
30
  still_on_hub: bool = False
31
 
32
  @classmethod
 
38
  config = data.get("config")
39
 
40
  # Precision
41
+ precision = utils.Precision.from_str(config.get("model_dtype"))
42
 
43
  # Get model and org
44
+ full_model = config.get("model_name", config.get("model_args", None))
45
+ org, model = full_model.split("/", 1) if "/" in full_model else (None, full_model)
46
 
47
+ if org:
 
 
 
 
 
 
48
  result_key = f"{org}_{model}_{precision.value.name}"
49
+ else:
50
+ result_key = f"{model}_{precision.value.name}"
51
 
52
+ still_on_hub, _, model_config = check_validity.is_model_on_hub(
53
+ full_model, config.get("model_sha", "main"), trust_remote_code=True,
54
+ test_tokenizer=False)
55
+
56
+ if model_config:
57
+ architecture = ";".join(getattr(model_config, "architectures", ["?"]))
58
+ else:
59
+ architecture = "?"
60
 
61
  # Extract results available in this file (some results are split in several files)
62
  results = {}
63
+ for task in utils.Tasks:
64
  task = task.value
65
 
66
  # We average all scores of a given metric (not all metrics are present in all files)
67
  accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark == k])
 
 
68
 
69
+ results[task.benchmark] = accs
 
70
 
71
  return self(
72
  eval_name=result_key,
 
74
  org=org,
75
  model=model,
76
  results=results,
77
+ precision=precision,
78
  revision= config.get("model_sha", ""),
79
  still_on_hub=still_on_hub,
80
  architecture=architecture
 
82
 
83
  def update_with_request_file(self, requests_path):
84
  """Finds the relevant request file for the current model and updates info with it"""
85
+ request_file = get_request_file_for_model(requests_path, self.full_model,
86
+ self.precision.value.name)
87
 
88
  try:
89
  with open(request_file, "r") as f:
90
  request = json.load(f)
91
+ self.model_type = utils.ModelType.from_str(request.get("model_type", ""))
92
+ self.weight_type = utils.WeightType[request.get("weight_type", "Original")]
93
  self.license = request.get("license", "?")
94
  self.likes = request.get("likes", 0)
95
  self.num_params = request.get("params", 0)
96
  self.date = request.get("submitted_time", "")
97
+ except FileNotFoundError:
98
+ print(f"Could not find request file for {self.org}/{self.model}")
99
+ except json.JSONDecodeError:
100
+ print(f"Error decoding JSON in request file for {self.org}/{self.model}")
101
 
102
  def to_dict(self):
103
  """Converts the Eval Result to a dict compatible with our dataframe display"""
104
+
105
  data_dict = {
106
  "eval_name": self.eval_name, # not a column, just a save name,
107
+ utils.AutoEvalColumn.precision.name: self.precision.value.name,
108
+ utils.AutoEvalColumn.model_type.name: self.model_type.value.name,
109
+ utils.AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol,
110
+ utils.AutoEvalColumn.weight_type.name: self.weight_type.value.name,
111
+ utils.AutoEvalColumn.architecture.name: self.architecture,
112
+ utils.AutoEvalColumn.model.name: formatting.make_clickable_model(self.full_model),
113
+ utils.AutoEvalColumn.dummy.name: self.full_model,
114
+ utils.AutoEvalColumn.revision.name: self.revision,
115
+ utils.AutoEvalColumn.license.name: self.license,
116
+ utils.AutoEvalColumn.likes.name: self.likes,
117
+ utils.AutoEvalColumn.params.name: self.num_params,
118
+ utils.AutoEvalColumn.still_on_hub.name: self.still_on_hub,
119
  }
120
 
121
+ for task in utils.Tasks:
122
  data_dict[task.value.col_name] = self.results[task.value.benchmark]
123
 
124
  return data_dict
 
149
  def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
150
  """From the path of the results folder root, extract all needed info for results"""
151
  model_result_filepaths = []
152
+ print("results_path", results_path)
153
  for root, _, files in os.walk(results_path):
154
  # We should only have json files in model results
155
+ print("file",files)
 
156
 
157
+ # if not files or any([not f.endswith(".json") for f in files]):
158
+
159
+ # continue
160
+ for f in files:
161
+ if f.endswith(".json"):
 
 
 
162
 
163
+ # Sort the files by date
164
+ # try:
165
+ # files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
166
+ # except dateutil.parser._parser.ParserError:
167
+ # files = [files[-1]]
168
+
169
+ model_result_filepaths.extend([os.path.join(root, f)])
170
+ print("model_result_filepaths", model_result_filepaths)
171
+ # exit()
172
  eval_results = {}
173
  for model_result_filepath in model_result_filepaths:
174
  # Creation of result
 
178
  # Store results of same eval together
179
  eval_name = eval_result.eval_name
180
  if eval_name in eval_results.keys():
181
+ eval_results[eval_name].results.update({k: v for k, v in
182
+ eval_result.results.items() if v is not None})
183
  else:
184
  eval_results[eval_name] = eval_result
185
 
src/populate.py CHANGED
@@ -3,27 +3,28 @@ import os
3
 
4
  import pandas as pd
5
 
6
- from src.display.formatting import has_no_nan_values, make_clickable_model
7
- from src.display.utils import AutoEvalColumn, EvalQueueColumn
8
- from src.leaderboard.read_evals import get_raw_eval_results
9
 
10
 
11
  def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
12
- """Creates a dataframe from all the individual experiment results"""
13
- raw_data = get_raw_eval_results(results_path, requests_path)
14
  all_data_json = [v.to_dict() for v in raw_data]
15
-
 
16
  df = pd.DataFrame.from_records(all_data_json)
17
- df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
 
 
18
  df = df[cols].round(decimals=2)
19
 
20
  # filter out if any of the benchmarks have not been produced
21
- df = df[has_no_nan_values(df, benchmark_cols)]
22
- return df
23
 
24
 
25
  def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
26
- """Creates the different dataframes for the evaluation queues requestes"""
27
  entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
28
  all_evals = []
29
 
@@ -33,8 +34,8 @@ def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
33
  with open(file_path) as fp:
34
  data = json.load(fp)
35
 
36
- data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
37
- data[EvalQueueColumn.revision.name] = data.get("revision", "main")
38
 
39
  all_evals.append(data)
40
  elif ".md" not in entry:
@@ -45,8 +46,8 @@ def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
45
  with open(file_path) as fp:
46
  data = json.load(fp)
47
 
48
- data[EvalQueueColumn.model.name] = make_clickable_model(data["model"])
49
- data[EvalQueueColumn.revision.name] = data.get("revision", "main")
50
  all_evals.append(data)
51
 
52
  pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
 
3
 
4
  import pandas as pd
5
 
6
+ import src.display.formatting as formatting
7
+ import src.display.utils as utils
8
+ import src.leaderboard.read_evals as read_evals
9
 
10
 
11
  def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame:
12
+ raw_data = read_evals.get_raw_eval_results(results_path, requests_path)
 
13
  all_data_json = [v.to_dict() for v in raw_data]
14
+ print(results_path, requests_path)
15
+ print(all_data_json)
16
  df = pd.DataFrame.from_records(all_data_json)
17
+ print(df)
18
+ # exit()
19
+ df = df.sort_values(by=[utils.AutoEvalColumn.hallucination_rate.name], ascending=True)
20
  df = df[cols].round(decimals=2)
21
 
22
  # filter out if any of the benchmarks have not been produced
23
+ df = df[formatting.has_no_nan_values(df, benchmark_cols)]
24
+ return raw_data, df
25
 
26
 
27
  def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]:
 
28
  entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")]
29
  all_evals = []
30
 
 
34
  with open(file_path) as fp:
35
  data = json.load(fp)
36
 
37
+ data[utils.EvalQueueColumn.model.name] = formatting.make_clickable_model(data["model"])
38
+ data[utils.EvalQueueColumn.revision.name] = data.get("revision", "main")
39
 
40
  all_evals.append(data)
41
  elif ".md" not in entry:
 
46
  with open(file_path) as fp:
47
  data = json.load(fp)
48
 
49
+ data[utils.EvalQueueColumn.model.name] = formatting.make_clickable_model(data["model"])
50
+ data[utils.EvalQueueColumn.revision.name] = data.get("revision", "main")
51
  all_evals.append(data)
52
 
53
  pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]]
src/submission/check_validity.py CHANGED
@@ -1,14 +1,12 @@
1
  import json
2
  import os
3
- import re
4
  from collections import defaultdict
5
- from datetime import datetime, timedelta, timezone
6
 
7
  import huggingface_hub
8
  from huggingface_hub import ModelCard
9
  from huggingface_hub.hf_api import ModelInfo
10
- from transformers import AutoConfig
11
- from transformers.models.auto.tokenization_auto import AutoTokenizer
12
 
13
  def check_model_card(repo_id: str) -> tuple[bool, str]:
14
  """Checks if the model card and license exist and have been filled"""
@@ -31,8 +29,8 @@ def check_model_card(repo_id: str) -> tuple[bool, str]:
31
 
32
  return True, ""
33
 
 
34
  def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False) -> tuple[bool, str]:
35
- """Checks if the model model_name is on the hub, and whether it (and its tokenizer) can be loaded with AutoClasses."""
36
  try:
37
  config = AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
38
  if test_tokenizer:
@@ -56,7 +54,8 @@ def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_rem
56
  )
57
 
58
  except Exception as e:
59
- return False, "was not found on hub!", None
 
60
 
61
 
62
  def get_model_size(model_info: ModelInfo, precision: str):
@@ -75,7 +74,6 @@ def get_model_arch(model_info: ModelInfo):
75
  return model_info.config.get("architectures", "Unknown")
76
 
77
  def already_submitted_models(requested_models_dir: str) -> set[str]:
78
- """Gather a list of already submitted models to avoid duplicates"""
79
  depth = 1
80
  file_names = []
81
  users_to_submission_dates = defaultdict(list)
 
1
  import json
2
  import os
 
3
  from collections import defaultdict
 
4
 
5
  import huggingface_hub
6
  from huggingface_hub import ModelCard
7
  from huggingface_hub.hf_api import ModelInfo
8
+ from transformers import AutoConfig, AutoTokenizer
9
+ from transformers.models.auto.tokenization_auto import tokenizer_class_from_name, get_tokenizer_config
10
 
11
  def check_model_card(repo_id: str) -> tuple[bool, str]:
12
  """Checks if the model card and license exist and have been filled"""
 
29
 
30
  return True, ""
31
 
32
+
33
  def is_model_on_hub(model_name: str, revision: str, token: str = None, trust_remote_code=False, test_tokenizer=False) -> tuple[bool, str]:
 
34
  try:
35
  config = AutoConfig.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
36
  if test_tokenizer:
 
54
  )
55
 
56
  except Exception as e:
57
+ return False, f"was not found on hub!: {e}", None
58
+
59
 
60
 
61
  def get_model_size(model_info: ModelInfo, precision: str):
 
74
  return model_info.config.get("architectures", "Unknown")
75
 
76
  def already_submitted_models(requested_models_dir: str) -> set[str]:
 
77
  depth = 1
78
  file_names = []
79
  users_to_submission_dates = defaultdict(list)
src/submission/submit.py CHANGED
@@ -2,14 +2,10 @@ import json
2
  import os
3
  from datetime import datetime, timezone
4
 
5
- from src.display.formatting import styled_error, styled_message, styled_warning
6
- from src.envs import API, EVAL_REQUESTS_PATH, TOKEN, QUEUE_REPO
7
- from src.submission.check_validity import (
8
- already_submitted_models,
9
- check_model_card,
10
- get_model_size,
11
- is_model_on_hub,
12
- )
13
 
14
  REQUESTED_MODELS = None
15
  USERS_TO_SUBMISSION_DATES = None
@@ -25,7 +21,7 @@ def add_new_eval(
25
  global REQUESTED_MODELS
26
  global USERS_TO_SUBMISSION_DATES
27
  if not REQUESTED_MODELS:
28
- REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = already_submitted_models(EVAL_REQUESTS_PATH)
29
 
30
  user_name = ""
31
  model_path = model
@@ -37,7 +33,7 @@ def add_new_eval(
37
  current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
38
 
39
  if model_type is None or model_type == "":
40
- return styled_error("Please select a model type.")
41
 
42
  # Does the model actually exist?
43
  if revision == "":
@@ -45,32 +41,32 @@ def add_new_eval(
45
 
46
  # Is the model on the hub?
47
  if weight_type in ["Delta", "Adapter"]:
48
- base_model_on_hub, error, _ = is_model_on_hub(model_name=base_model, revision=revision, token=TOKEN, test_tokenizer=True)
49
  if not base_model_on_hub:
50
- return styled_error(f'Base model "{base_model}" {error}')
51
 
52
  if not weight_type == "Adapter":
53
- model_on_hub, error, _ = is_model_on_hub(model_name=model, revision=revision, token=TOKEN, test_tokenizer=True)
54
  if not model_on_hub:
55
- return styled_error(f'Model "{model}" {error}')
56
 
57
  # Is the model info correctly filled?
58
  try:
59
- model_info = API.model_info(repo_id=model, revision=revision)
60
  except Exception:
61
- return styled_error("Could not get your model information. Please fill it up properly.")
62
 
63
- model_size = get_model_size(model_info=model_info, precision=precision)
64
 
65
  # Were the model card and license filled?
66
  try:
67
  license = model_info.cardData["license"]
68
  except Exception:
69
- return styled_error("Please select a license for your model")
70
 
71
- modelcard_OK, error_msg = check_model_card(model)
72
  if not modelcard_OK:
73
- return styled_error(error_msg)
74
 
75
  # Seems good, creating the eval
76
  print("Adding new eval")
@@ -87,15 +83,15 @@ def add_new_eval(
87
  "likes": model_info.likes,
88
  "params": model_size,
89
  "license": license,
90
- "private": False,
91
  }
92
 
93
  # Check for duplicate submission
94
  if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
95
- return styled_warning("This model has been already submitted.")
96
 
97
  print("Creating eval file")
98
- OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
 
99
  os.makedirs(OUT_DIR, exist_ok=True)
100
  out_path = f"{OUT_DIR}/{model_path}_eval_request_False_{precision}_{weight_type}.json"
101
 
@@ -103,10 +99,10 @@ def add_new_eval(
103
  f.write(json.dumps(eval_entry))
104
 
105
  print("Uploading eval file")
106
- API.upload_file(
107
  path_or_fileobj=out_path,
108
  path_in_repo=out_path.split("eval-queue/")[1],
109
- repo_id=QUEUE_REPO,
110
  repo_type="dataset",
111
  commit_message=f"Add {model} to eval queue",
112
  )
@@ -114,6 +110,6 @@ def add_new_eval(
114
  # Remove the local file
115
  os.remove(out_path)
116
 
117
- return styled_message(
118
  "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
119
  )
 
2
  import os
3
  from datetime import datetime, timezone
4
 
5
+ import src.display.formatting as formatting
6
+ import src.envs as envs
7
+ import src.submission.check_validity as check_validity
8
+
 
 
 
 
9
 
10
  REQUESTED_MODELS = None
11
  USERS_TO_SUBMISSION_DATES = None
 
21
  global REQUESTED_MODELS
22
  global USERS_TO_SUBMISSION_DATES
23
  if not REQUESTED_MODELS:
24
+ REQUESTED_MODELS, USERS_TO_SUBMISSION_DATES = check_validity.already_submitted_models(envs.EVAL_REQUESTS_PATH)
25
 
26
  user_name = ""
27
  model_path = model
 
33
  current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
34
 
35
  if model_type is None or model_type == "":
36
+ return formatting.styled_error("Please select a model type.")
37
 
38
  # Does the model actually exist?
39
  if revision == "":
 
41
 
42
  # Is the model on the hub?
43
  if weight_type in ["Delta", "Adapter"]:
44
+ base_model_on_hub, error, _ = check_validity.is_model_on_hub(model_name=base_model, revision=revision, token=envs.TOKEN, test_tokenizer=True)
45
  if not base_model_on_hub:
46
+ return formatting.styled_error(f'Base model "{base_model}" {error}')
47
 
48
  if not weight_type == "Adapter":
49
+ model_on_hub, error, _ = check_validity.is_model_on_hub(model_name=model, revision=revision, test_tokenizer=True)
50
  if not model_on_hub:
51
+ return formatting.styled_error(f'Model "{model}" {error}')
52
 
53
  # Is the model info correctly filled?
54
  try:
55
+ model_info = envs.API.model_info(repo_id=model, revision=revision)
56
  except Exception:
57
+ return formatting.styled_error("Could not get your model information. Please fill it up properly.")
58
 
59
+ model_size = check_validity.get_model_size(model_info=model_info, precision=precision)
60
 
61
  # Were the model card and license filled?
62
  try:
63
  license = model_info.cardData["license"]
64
  except Exception:
65
+ return formatting.styled_error("Please select a license for your model")
66
 
67
+ modelcard_OK, error_msg = check_validity.check_model_card(model)
68
  if not modelcard_OK:
69
+ return formatting.styled_error(error_msg)
70
 
71
  # Seems good, creating the eval
72
  print("Adding new eval")
 
83
  "likes": model_info.likes,
84
  "params": model_size,
85
  "license": license,
 
86
  }
87
 
88
  # Check for duplicate submission
89
  if f"{model}_{revision}_{precision}" in REQUESTED_MODELS:
90
+ return formatting.styled_warning("This model has been already submitted.")
91
 
92
  print("Creating eval file")
93
+
94
+ OUT_DIR = f"{envs.EVAL_REQUESTS_PATH}/{user_name}"
95
  os.makedirs(OUT_DIR, exist_ok=True)
96
  out_path = f"{OUT_DIR}/{model_path}_eval_request_False_{precision}_{weight_type}.json"
97
 
 
99
  f.write(json.dumps(eval_entry))
100
 
101
  print("Uploading eval file")
102
+ envs.API.upload_file(
103
  path_or_fileobj=out_path,
104
  path_in_repo=out_path.split("eval-queue/")[1],
105
+ repo_id=envs.QUEUE_REPO,
106
  repo_type="dataset",
107
  commit_message=f"Add {model} to eval queue",
108
  )
 
110
  # Remove the local file
111
  os.remove(out_path)
112
 
113
+ return formatting.styled_message(
114
  "Your request has been submitted to the evaluation queue!\nPlease wait for up to an hour for the model to show in the PENDING list."
115
  )