ToluClassics commited on
Commit
c25e6bb
1 Parent(s): 6364517

mockup eval

Browse files
Makefile DELETED
@@ -1,13 +0,0 @@
1
- .PHONY: style format
2
-
3
-
4
- style:
5
- python -m black --line-length 119 .
6
- python -m isort .
7
- ruff check --fix .
8
-
9
-
10
- quality:
11
- python -m black --check --line-length 119 .
12
- python -m isort --check-only .
13
- ruff check .
 
 
 
 
 
 
 
 
 
 
 
 
 
 
app.py CHANGED
@@ -1,204 +1,162 @@
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(
111
- f"✅ Finished Evaluations ({len(finished_eval_queue_df)})",
112
- open=False,
113
- ):
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(
122
- f"🔄 Running Evaluation Queue ({len(running_eval_queue_df)})",
123
- open=False,
124
- ):
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
-
133
- with gr.Accordion(
134
- f"⏳ Pending Evaluation Queue ({len(pending_eval_queue_df)})",
135
- open=False,
136
- ):
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():
145
- gr.Markdown("# ✉️✨ Submit your model here!", elem_classes="markdown-text")
146
-
147
- with gr.Row():
148
- with gr.Column():
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,
156
- interactive=True,
157
- )
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",
172
- interactive=True,
173
- )
174
- base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)")
175
-
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,
183
- revision_name_textbox,
184
- precision,
185
- weight_type,
186
- model_type,
187
- ],
188
- submission_result,
189
- )
190
-
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,
199
- )
200
-
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 os
2
+ import json
3
+ import glob
4
+ from collections import defaultdict
5
  import pandas as pd
6
+ import gradio as gr
7
+ from content import *
8
+ from css import *
9
+ import glob
10
+
11
+ AFRIMMLU_DIRECT = "afrimmlu_direct"
12
+ AFRIMMLU_TRANSLATE = "afrimmlu_translate"
13
+ AFRIXNLI_DIRECT = "afrixnli_direct"
14
+ AFRIXNLI_TRANSLATE = "afrixnli_translate"
15
+
16
+
17
+ BENCHMARKS = [AFRIMMLU_DIRECT, AFRIMMLU_TRANSLATE, AFRIXNLI_DIRECT, AFRIXNLI_TRANSLATE]
18
+
19
+ METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]
20
+
21
+ LANGS = ['amh', 'eng', 'ewe', 'fra', 'hau', 'ibo', 'kin', 'lin', 'lug', 'orm', 'sna', 'sot', 'swa', 'twi', 'wol', 'xho', 'yor', 'zul']
22
+
23
+ LANG_NAME = {
24
+ 'amh': 'Amharic',
25
+ 'eng': 'English',
26
+ 'ewe': 'Ewe',
27
+ 'fra': 'French',
28
+ 'hau': 'Hausa',
29
+ 'ibo': 'Igbo',
30
+ 'kin': 'Kinyarwanda',
31
+ 'lin': 'Lingala',
32
+ 'lug': 'Luganda',
33
+ 'orm': 'Oromo',
34
+ 'sna': 'Shona',
35
+ 'sot': 'Sotho',
36
+ 'swa': 'Swahili',
37
+ 'twi': 'Twi',
38
+ 'wol': 'Wolof',
39
+ 'xho': 'Xhosa',
40
+ 'yor': 'Yoruba',
41
+ 'zul': 'Zulu'
42
+ }
43
+
44
+
45
+ def collect_results():
46
+ performance_dict = defaultdict(dict)
47
+ pretrained_models = set()
48
+ for file in glob.glob('evals/*/*.json'):
49
+ with open(file, 'r') as f:
50
+ data = json.load(f)
51
+ if 'results' not in data:
52
+ continue
53
+ if 'config' not in data:
54
+ continue
55
+ results = data['results']
56
+ config = data['config']
57
+ if 'model_args' not in config:
58
+ continue
59
+
60
+ model_args = config['model_args'].split(',')
61
+ pretrained = [x for x in model_args if x.startswith('pretrained=')]
62
+ if len(pretrained) != 1:
63
+ continue
64
+ pretrained = pretrained[0].split('=')[1]
65
+ pretrained = pretrained.split('/')[-1]
66
+ pretrained_models.add(pretrained)
67
+
68
+ for lang_task, perfs in results.items():
69
+ task, lang = lang_task.split('_')
70
+ assert task in BENCHMARKS
71
+
72
+ if lang and task:
73
+ metric = METRICS[BENCHMARKS.index(task)]
74
+ p = round(perfs[metric] * 100, 1)
75
+ performance_dict[(pretrained, lang)][task] = p
76
+ return performance_dict, pretrained_models
77
+
78
+
79
+ def get_leaderboard_df(performance_dict, pretrained_models):
80
+ df = list()
81
+ for (pretrained, lang), perfs in performance_dict.items():
82
+ lang_name = LANG_NAME[lang]
83
+ afrimmlu_direct_perf = perfs.get(AFRIMMLU_DIRECT, 0.0)
84
+ afrimmlu_translate_perf = perfs.get(AFRIMMLU_TRANSLATE, 0.0)
85
+ afrixnli_direct_perf = perfs.get(AFRIXNLI_DIRECT, 0.0)
86
+ afrixnli_translate_perf = perfs.get(AFRIXNLI_TRANSLATE, 0.0)
87
+
88
+ if afrimmlu_direct_perf * afrimmlu_translate_perf * afrixnli_direct_perf * afrixnli_translate_perf == 0:
89
+ continue
90
+
91
+ avg = round((afrimmlu_direct_perf + afrimmlu_translate_perf + afrixnli_direct_perf + afrixnli_translate_perf) / 4, 1)
92
+ notes = ' '.join([pretrained, lang_name])
93
+ row = [pretrained, lang_name, lang, avg, afrimmlu_direct_perf, afrimmlu_translate_perf, afrixnli_direct_perf, afrixnli_translate_perf, notes]
94
+ df.append(row)
95
+
96
+ df = pd.DataFrame.from_records(df, columns=COLS)
97
+ df = df.sort_values(by=[LANG_COL, AVERAGE_COL], ascending=False)
98
+ df = df[COLS]
99
+
100
+ return df
101
+
102
+
103
+ def search_table(df, query):
104
+ filtered_df = df[df[NOTES_COL].str.contains(query, case=False)]
105
+ return filtered_df
106
+
107
+
108
+
109
+ MODEL_COL = "Model"
110
+ LANG_COL = "Language"
111
+ CODE_COL = "Code"
112
+ AVERAGE_COL = "Average"
113
+ AFRIMMLU_DIRECT_COL = "AfriMMLU Direct (0-Shot)"
114
+ AFRIMMLU_TRANSLATE_COL = "AfriMMLU Translate (0-Shot)"
115
+ AFRIXNLI_DIRECT_COL = "AfriXNLI Direct (0-Shot)"
116
+ AFRIXNLI_TRANSLATE_COL = "AfriXNLI Translate (0-Shot)"
117
+ NOTES_COL = "Notes" # For search only
118
+
119
+ COLS = [MODEL_COL, LANG_COL, CODE_COL, AVERAGE_COL, AFRIMMLU_DIRECT_COL, AFRIMMLU_TRANSLATE_COL, AFRIXNLI_DIRECT_COL, AFRIXNLI_TRANSLATE_COL, NOTES_COL]
120
+ TYPES = ["str", "str", "str", "number", "number", "number", "number", "number", "str"]
121
+
122
+ args = collect_results()
123
+ original_df = get_leaderboard_df(*args)
124
+
125
+ demo = gr.Blocks(css=CUSTOM_CSS)
126
  with demo:
127
  gr.HTML(TITLE)
128
+ gr.Markdown(INTRO_TEXT, elem_classes="markdown-text")
129
+ gr.Markdown(HOW_TO, elem_classes="markdown-text")
130
+
131
+ with gr.Group():
132
+ search_bar = gr.Textbox(
133
+ placeholder="Search models and languages...", show_label=False, elem_id="search-bar"
134
+ )
135
+
136
+ leaderboard_table = gr.components.Dataframe(
137
+ value=original_df,
138
+ headers=COLS,
139
+ datatype=TYPES,
140
+ # max_rows=5,
141
+ elem_id="leaderboard-table",
142
+ )
143
+
144
+ # # Dummy leaderboard for handling the case when the user uses backspace key
145
+ hidden_leaderboard_table_for_search = gr.components.Dataframe(
146
+ value=original_df,
147
+ headers=COLS,
148
+ datatype=TYPES,
149
+ # max_rows=5,
150
+ visible=False
151
+ )
152
+
153
+ search_bar.change(
154
+ search_table,
155
+ [hidden_leaderboard_table_for_search, search_bar],
156
+ leaderboard_table,
157
+ )
158
+
159
+ gr.Markdown(CREDIT, elem_classes="markdown-text")
160
+ gr.Markdown(CITATION, elem_classes="markdown-text")
161
+
162
+ demo.launch()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
content.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ TITLE = '<h1 align="center" id="space-title">African Languages LLM Eval Leaderboard</h1>'
2
+
3
+ INTRO_TEXT = f"""
4
+ ## About
5
+ This leaderboard tracks progress and ranks performance of large language models (LLMs) on African languages.
6
+
7
+ We currently evaluate models over the following benchmarks:
8
+
9
+ - <a href="https://arxiv.org/abs/2406.03368" target="_blank"> Iroko Bench </a>
10
+ - AfriMMLU (Direct and Translate) (0-shot)
11
+ - AfriXNLI (Direct and Translate) (0-shot)
12
+
13
+ - <a href="https://aclanthology.org/2024.acl-long.44" target="_blank"> BeleBele </a>
14
+ - <a href="https://arxiv.org/abs/2305.06897" target="_blank"> AfriQa </a>
15
+ """
16
+
17
+ HOW_TO = f"""
18
+ ## How to list your model performance on this leaderboard:
19
+ Run the evaluation of your model using this repo: <a href="https://github.com/The-African-Research-Collective/jubilant-sniffle" target="_blank">https://github.com/The-African-Research-Collective/jubilant-sniffle</a>.
20
+ And then, push the evaluation log and make a pull request.
21
+ """
22
+
23
+ CREDIT = f"""
24
+ ## Credit
25
+ To make this website, we use the following resources:
26
+ - Datasets (IrokoBench, AfriQA, BeleBele)
27
+ - Evaluation code (EleutherAI's lm_evaluation_harness repo)
28
+ - Multilingual LLM Leaderboard code (UONLP's Open Multilingual LLM Evaluation Leaderboard)
29
+ """
30
+
31
+
32
+ CITATION = f"""
33
+ ## Citation
34
+ ```
35
+ @misc{{tarescollmbenchmark,
36
+ author = {{tbd...}},
37
+ title={{tbd...}},
38
+ year={{2024}}
39
+ }}
40
+ ```
41
+ """
css.py ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ CUSTOM_CSS= """
2
+ /* Hides the final column */
3
+ table td:last-child,
4
+ table th:last-child {
5
+ display: none;
6
+ }
7
+ # table td:first-child,
8
+ # table th:first-child {
9
+ # max-width: 400px;
10
+ # overflow: auto;
11
+ # white-space: nowrap;
12
+ # }
13
+ """
evals/afrimmlu_direct/afrimmlu_direct_amh-Meta-Llama-3-8B-Instruct.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "results": {
3
+ "afrimmlu_direct_amh": {
4
+ "acc": 0.2634730538922156,
5
+ "acc_stderr": 0.012889646336321774,
6
+ "acc_norm": 0.31394354148845166,
7
+ "acc_norm_stderr": 0.013579515768185788
8
+ }
9
+ },
10
+ "versions": {
11
+ "afrimmlu_direct_amh": 0
12
+ },
13
+ "config": {
14
+ "model": "hf",
15
+ "model_args": "pretrained=meta-llama/Meta-Llama-3-8B-Instruct",
16
+ "batch_size": 8,
17
+ "device": "cuda",
18
+ "no_cache": false,
19
+ "limit": null,
20
+ "bootstrap_iters": 100000,
21
+ "description_dict": {}
22
+ }
23
+ }
evals/afrimmlu_translate/afrimmlu_translate_amh-Meta-Llama-3-8B-Instruct.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "results": {
3
+ "afrimmlu_translate_amh": {
4
+ "acc": 0.2634730538922156,
5
+ "acc_stderr": 0.012889646336321774,
6
+ "acc_norm": 0.31394354148845166,
7
+ "acc_norm_stderr": 0.013579515768185788
8
+ }
9
+ },
10
+ "versions": {
11
+ "afrimmlu_translate_amh": 0
12
+ },
13
+ "config": {
14
+ "model": "hf",
15
+ "model_args": "pretrained=meta-llama/Meta-Llama-3-8B-Instruct",
16
+ "batch_size": 8,
17
+ "device": "cuda",
18
+ "no_cache": false,
19
+ "limit": null,
20
+ "bootstrap_iters": 100000,
21
+ "description_dict": {}
22
+ }
23
+ }
evals/afrixnli_direct/afrixnli_direct_amh-Meta-Llama-3-8B-Instruct.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "results": {
3
+ "afrixnli_direct_amh": {
4
+ "acc": 0.2634730538922156,
5
+ "acc_stderr": 0.012889646336321774,
6
+ "acc_norm": 0.31394354148845166,
7
+ "acc_norm_stderr": 0.013579515768185788
8
+ }
9
+ },
10
+ "versions": {
11
+ "afrixnli_direct_amh": 0
12
+ },
13
+ "config": {
14
+ "model": "hf",
15
+ "model_args": "pretrained=meta-llama/Meta-Llama-3-8B-Instruct",
16
+ "batch_size": 8,
17
+ "device": "cuda",
18
+ "no_cache": false,
19
+ "limit": null,
20
+ "bootstrap_iters": 100000,
21
+ "description_dict": {}
22
+ }
23
+ }
evals/afrixnli_translate/afrixnli_translate_amh-Meta-Llama-3-8B-Instruct.json ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "results": {
3
+ "afrixnli_translate_amh": {
4
+ "acc": 0.2634730538922156,
5
+ "acc_stderr": 0.012889646336321774,
6
+ "acc_norm": 0.31394354148845166,
7
+ "acc_norm_stderr": 0.013579515768185788
8
+ }
9
+ },
10
+ "versions": {
11
+ "afrixnli_translate_amh": 0
12
+ },
13
+ "config": {
14
+ "model": "hf",
15
+ "model_args": "pretrained=meta-llama/Meta-Llama-3-8B-Instruct",
16
+ "batch_size": 8,
17
+ "device": "cuda",
18
+ "no_cache": false,
19
+ "limit": null,
20
+ "bootstrap_iters": 100000,
21
+ "description_dict": {}
22
+ }
23
+ }
src/about.py DELETED
@@ -1,72 +0,0 @@
1
- from dataclasses import dataclass
2
- from enum import Enum
3
-
4
- @dataclass
5
- class Task:
6
- benchmark: str
7
- metric: str
8
- col_name: str
9
-
10
-
11
- # Select your tasks here
12
- # ---------------------------------------------------
13
- class Tasks(Enum):
14
- # task_key in the json file, metric_key in the json file, name to display in the leaderboard
15
- task0 = Task("anli_r1", "acc", "ANLI")
16
- task1 = Task("logiqa", "acc_norm", "LogiQA")
17
-
18
- NUM_FEWSHOT = 0 # Change with your few shot
19
- # ---------------------------------------------------
20
-
21
-
22
-
23
- # Your leaderboard name
24
- TITLE = """<h1 align="center" id="space-title">Demo leaderboard</h1>"""
25
-
26
- # What does your leaderboard evaluate?
27
- INTRODUCTION_TEXT = """
28
- Intro text
29
- """
30
-
31
- # Which evaluations are you running? how can people reproduce what you have?
32
- LLM_BENCHMARKS_TEXT = f"""
33
- ## How it works
34
-
35
- ## Reproducibility
36
- To reproduce our results, here is the commands you can run:
37
-
38
- """
39
-
40
- EVALUATION_QUEUE_TEXT = """
41
- ## Some good practices before submitting a model
42
-
43
- ### 1) Make sure you can load your model and tokenizer using AutoClasses:
44
- ```python
45
- from transformers import AutoConfig, AutoModel, AutoTokenizer
46
- config = AutoConfig.from_pretrained("your model name", revision=revision)
47
- model = AutoModel.from_pretrained("your model name", revision=revision)
48
- tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
49
- ```
50
- If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded.
51
-
52
- Note: make sure your model is public!
53
- Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!
54
-
55
- ### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
56
- It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!
57
-
58
- ### 3) Make sure your model has an open license!
59
- This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗
60
-
61
- ### 4) Fill up your model card
62
- When we add extra information about models to the leaderboard, it will be automatically taken from the model card
63
-
64
- ## In case of model failure
65
- If your model is displayed in the `FAILED` category, its execution stopped.
66
- Make sure you have followed the above steps first.
67
- If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).
68
- """
69
-
70
- CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
71
- CITATION_BUTTON_TEXT = r"""
72
- """
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/display/css_html_js.py DELETED
@@ -1,105 +0,0 @@
1
- custom_css = """
2
-
3
- .markdown-text {
4
- font-size: 16px !important;
5
- }
6
-
7
- #models-to-add-text {
8
- font-size: 18px !important;
9
- }
10
-
11
- #citation-button span {
12
- font-size: 16px !important;
13
- }
14
-
15
- #citation-button textarea {
16
- font-size: 16px !important;
17
- }
18
-
19
- #citation-button > label > button {
20
- margin: 6px;
21
- transform: scale(1.3);
22
- }
23
-
24
- #leaderboard-table {
25
- margin-top: 15px
26
- }
27
-
28
- #leaderboard-table-lite {
29
- margin-top: 15px
30
- }
31
-
32
- #search-bar-table-box > div:first-child {
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 {
44
- max-width: 400px;
45
- overflow: auto;
46
- white-space: nowrap;
47
- }
48
-
49
- .tab-buttons button {
50
- font-size: 20px;
51
- }
52
-
53
- #scale-logo {
54
- border-style: none !important;
55
- box-shadow: none;
56
- display: block;
57
- margin-left: auto;
58
- margin-right: auto;
59
- max-width: 600px;
60
- }
61
-
62
- #scale-logo .download {
63
- display: none;
64
- }
65
- #filter_type{
66
- border: 0;
67
- padding-left: 0;
68
- padding-top: 0;
69
- }
70
- #filter_type label {
71
- display: flex;
72
- }
73
- #filter_type label > span{
74
- margin-top: var(--spacing-lg);
75
- margin-right: 0.5em;
76
- }
77
- #filter_type label > .wrap{
78
- width: 103px;
79
- }
80
- #filter_type label > .wrap .wrap-inner{
81
- padding: 2px;
82
- }
83
- #filter_type label > .wrap .wrap-inner input{
84
- width: 1px
85
- }
86
- #filter-columns-type{
87
- border:0;
88
- padding:0.5;
89
- }
90
- #filter-columns-size{
91
- border:0;
92
- padding:0.5;
93
- }
94
- #box-filter > .form{
95
- border: 0
96
- }
97
- """
98
-
99
- get_window_url_params = """
100
- function(url_params) {
101
- const params = new URLSearchParams(window.location.search);
102
- url_params = Object.fromEntries(params);
103
- return url_params;
104
- }
105
- """
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/display/formatting.py DELETED
@@ -1,27 +0,0 @@
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
-
4
-
5
- def make_clickable_model(model_name):
6
- link = f"https://huggingface.co/{model_name}"
7
- return model_hyperlink(link, model_name)
8
-
9
-
10
- def styled_error(error):
11
- return f"<p style='color: red; font-size: 20px; text-align: center;'>{error}</p>"
12
-
13
-
14
- def styled_warning(warn):
15
- return f"<p style='color: orange; font-size: 20px; text-align: center;'>{warn}</p>"
16
-
17
-
18
- def styled_message(message):
19
- return f"<p style='color: green; font-size: 20px; text-align: center;'>{message}</p>"
20
-
21
-
22
- def has_no_nan_values(df, columns):
23
- return df[columns].notna().all(axis=1)
24
-
25
-
26
- def has_nan_values(df, columns):
27
- return df[columns].isna().any(axis=1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/display/utils.py DELETED
@@ -1,110 +0,0 @@
1
- from dataclasses import dataclass, make_dataclass
2
- 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:] != "__"]
10
-
11
-
12
- # These classes are for user facing column names,
13
- # to avoid having to change them all around the code
14
- # when a modif is needed
15
- @dataclass
16
- class ColumnContent:
17
- name: str
18
- type: str
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)])
35
- auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
36
- auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
37
- auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
38
- auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
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)
45
-
46
- ## For the queue columns in the submission tab
47
- @dataclass(frozen=True)
48
- class EvalQueueColumn: # Queue column
49
- model = ColumnContent("model", "markdown", True)
50
- revision = ColumnContent("revision", "str", True)
51
- private = ColumnContent("private", "bool", True)
52
- precision = ColumnContent("precision", "str", True)
53
- weight_type = ColumnContent("weight_type", "str", "Original")
54
- status = ColumnContent("status", "str", True)
55
-
56
- ## All the model information that we might need
57
- @dataclass
58
- class ModelDetails:
59
- name: str
60
- display_name: str = ""
61
- symbol: str = "" # emoji
62
-
63
-
64
- class ModelType(Enum):
65
- PT = ModelDetails(name="pretrained", symbol="🟢")
66
- FT = ModelDetails(name="fine-tuned", symbol="🔶")
67
- IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
68
- RL = ModelDetails(name="RL-tuned", symbol="🟦")
69
- Unknown = ModelDetails(name="", symbol="?")
70
-
71
- def to_str(self, separator=" "):
72
- return f"{self.value.symbol}{separator}{self.value.name}"
73
-
74
- @staticmethod
75
- def from_str(type):
76
- if "fine-tuned" in type or "🔶" in type:
77
- return ModelType.FT
78
- if "pretrained" in type or "🟢" in type:
79
- return ModelType.PT
80
- if "RL-tuned" in type or "🟦" in type:
81
- return ModelType.RL
82
- if "instruction-tuned" in type or "⭕" in type:
83
- return ModelType.IFT
84
- return ModelType.Unknown
85
-
86
- class WeightType(Enum):
87
- Adapter = ModelDetails("Adapter")
88
- Original = ModelDetails("Original")
89
- Delta = ModelDetails("Delta")
90
-
91
- class Precision(Enum):
92
- float16 = ModelDetails("float16")
93
- bfloat16 = ModelDetails("bfloat16")
94
- Unknown = ModelDetails("?")
95
-
96
- def from_str(precision):
97
- if precision in ["torch.float16", "float16"]:
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
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/envs.py DELETED
@@ -1,25 +0,0 @@
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
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)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/leaderboard/read_evals.py DELETED
@@ -1,196 +0,0 @@
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
36
- def init_from_json_file(self, json_filepath):
37
- """Inits the result from the specific model result file"""
38
- with open(json_filepath) as fp:
39
- data = json.load(fp)
40
-
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,
84
- full_model=full_model,
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
92
- )
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
133
-
134
-
135
- def get_request_file_for_model(requests_path, model_name, precision):
136
- """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
137
- request_files = os.path.join(
138
- requests_path,
139
- f"{model_name}_eval_request_*.json",
140
- )
141
- request_files = glob.glob(request_files)
142
-
143
- # Select correct request file (precision)
144
- request_file = ""
145
- request_files = sorted(request_files, reverse=True)
146
- for tmp_request_file in request_files:
147
- with open(tmp_request_file, "r") as f:
148
- req_content = json.load(f)
149
- if (
150
- req_content["status"] in ["FINISHED"]
151
- and req_content["precision"] == precision.split(".")[-1]
152
- ):
153
- request_file = tmp_request_file
154
- return request_file
155
-
156
-
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
178
- eval_result = EvalResult.init_from_json_file(model_result_filepath)
179
- eval_result.update_with_request_file(requests_path)
180
-
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
-
188
- results = []
189
- for v in eval_results.values():
190
- try:
191
- v.to_dict() # we test if the dict version is complete
192
- results.append(v)
193
- except KeyError: # not all eval values present
194
- continue
195
-
196
- return results
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/populate.py DELETED
@@ -1,58 +0,0 @@
1
- import json
2
- 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
-
30
- for entry in entries:
31
- if ".json" in entry:
32
- file_path = os.path.join(save_path, entry)
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:
41
- # this is a folder
42
- sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if not e.startswith(".")]
43
- for sub_entry in sub_entries:
44
- file_path = os.path.join(save_path, entry, sub_entry)
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"]]
53
- running_list = [e for e in all_evals if e["status"] == "RUNNING"]
54
- finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"]
55
- df_pending = pd.DataFrame.from_records(pending_list, columns=cols)
56
- df_running = pd.DataFrame.from_records(running_list, columns=cols)
57
- df_finished = pd.DataFrame.from_records(finished_list, columns=cols)
58
- return df_finished[cols], df_running[cols], df_pending[cols]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/submission/check_validity.py DELETED
@@ -1,99 +0,0 @@
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"""
15
- try:
16
- card = ModelCard.load(repo_id)
17
- except huggingface_hub.utils.EntryNotFoundError:
18
- return False, "Please add a model card to your model to explain how you trained/fine-tuned it."
19
-
20
- # Enforce license metadata
21
- if card.data.license is None:
22
- if not ("license_name" in card.data and "license_link" in card.data):
23
- return False, (
24
- "License not found. Please add a license to your model card using the `license` metadata or a"
25
- " `license_name`/`license_link` pair."
26
- )
27
-
28
- # Enforce card content
29
- if len(card.text) < 200:
30
- return False, "Please add a description to your model card, it is too short."
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:
39
- try:
40
- tk = AutoTokenizer.from_pretrained(model_name, revision=revision, trust_remote_code=trust_remote_code, token=token)
41
- except ValueError as e:
42
- return (
43
- False,
44
- f"uses a tokenizer which is not in a transformers release: {e}",
45
- None
46
- )
47
- except Exception as e:
48
- return (False, "'s tokenizer cannot be loaded. Is your tokenizer class in a stable transformers release, and correctly configured?", None)
49
- return True, None, config
50
-
51
- except ValueError:
52
- return (
53
- False,
54
- "needs to be launched with `trust_remote_code=True`. For safety reason, we do not allow these models to be automatically submitted to the leaderboard.",
55
- None
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):
63
- """Gets the model size from the configuration, or the model name if the configuration does not contain the information."""
64
- try:
65
- model_size = round(model_info.safetensors["total"] / 1e9, 3)
66
- except (AttributeError, TypeError):
67
- return 0 # Unknown model sizes are indicated as 0, see NUMERIC_INTERVALS in app.py
68
-
69
- size_factor = 8 if (precision == "GPTQ" or "gptq" in model_info.modelId.lower()) else 1
70
- model_size = size_factor * model_size
71
- return model_size
72
-
73
- def get_model_arch(model_info: ModelInfo):
74
- """Gets the model architecture from the configuration"""
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)
82
-
83
- for root, _, files in os.walk(requested_models_dir):
84
- current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
85
- if current_depth == depth:
86
- for file in files:
87
- if not file.endswith(".json"):
88
- continue
89
- with open(os.path.join(root, file), "r") as f:
90
- info = json.load(f)
91
- file_names.append(f"{info['model']}_{info['revision']}_{info['precision']}")
92
-
93
- # Select organisation
94
- if info["model"].count("/") == 0 or "submitted_time" not in info:
95
- continue
96
- organisation, _ = info["model"].split("/")
97
- users_to_submission_dates[organisation].append(info["submitted_time"])
98
-
99
- return set(file_names), users_to_submission_dates
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
src/submission/submit.py DELETED
@@ -1,119 +0,0 @@
1
- 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
16
-
17
- def add_new_eval(
18
- model: str,
19
- base_model: str,
20
- revision: str,
21
- precision: str,
22
- weight_type: str,
23
- model_type: str,
24
- ):
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
32
- if "/" in model:
33
- user_name = model.split("/")[0]
34
- model_path = model.split("/")[1]
35
-
36
- precision = precision.split(" ")[0]
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 == "":
44
- revision = "main"
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")
77
-
78
- eval_entry = {
79
- "model": model,
80
- "base_model": base_model,
81
- "revision": revision,
82
- "precision": precision,
83
- "weight_type": weight_type,
84
- "status": "PENDING",
85
- "submitted_time": current_time,
86
- "model_type": model_type,
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
-
102
- with open(out_path, "w") as f:
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
- )
113
-
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
- )