File size: 10,836 Bytes
e2e6875
25557b5
 
 
 
 
 
 
 
05c90f4
 
 
 
 
6679087
 
 
 
 
 
25557b5
7379857
 
38f4369
ddc25db
 
 
 
 
 
 
 
 
7379857
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38f4369
25557b5
 
 
c660995
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25557b5
 
3caeacd
 
 
 
05c90f4
c660995
25557b5
05c90f4
 
 
023a289
3caeacd
 
05c90f4
 
6679087
e29ab28
 
6679087
 
 
99aea78
 
 
 
6679087
 
81f1dd1
0d4c659
6679087
 
0d4c659
6679087
 
 
 
 
 
0d4c659
 
6679087
0d4c659
6679087
 
 
a0691fa
0d4c659
81f1dd1
6679087
 
 
3782698
5b4c5f8
 
 
 
 
 
 
25557b5
 
3caeacd
7379857
 
 
c8b695a
7379857
 
 
c8b695a
 
 
 
 
 
 
7379857
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
99aea78
 
 
 
eec78c0
 
 
 
7379857
eec78c0
7379857
 
 
 
 
 
 
 
e611814
 
 
 
 
023a289
e611814
 
 
023a289
6679087
e611814
023a289
6679087
e611814
 
2436603
6679087
05c90f4
3caeacd
 
 
 
 
 
 
 
 
 
 
 
7379857
3caeacd
 
 
 
 
 
 
7379857
3caeacd
7379857
 
 
 
c8b695a
7379857
 
 
 
25557b5
3caeacd
 
 
 
 
99aea78
 
 
6679087
 
 
 
5b4c5f8
3caeacd
5b4c5f8
ddc25db
 
6679087
 
 
7379857
 
3caeacd
 
 
 
 
c8b695a
 
 
 
 
 
7379857
99aea78
 
 
7379857
 
eec78c0
7379857
 
 
 
eec78c0
7379857
e611814
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
import io
import json

import gradio as gr
import pandas as pd
from huggingface_hub import HfFileSystem


RESULTS_DATASET_ID = "datasets/open-llm-leaderboard/results"
EXCLUDED_KEYS =  {
    "pretty_env_info",
    "chat_template",
    "group_subtasks",
}
# EXCLUDED_RESULTS_KEYS = {
#     "leaderboard",
# }
# EXCLUDED_RESULTS_LEADERBOARDS_KEYS = {
#     "alias",
# }

DETAILS_DATASET_ID = "datasets/open-llm-leaderboard/{model_name_sanitized}-details"
DETAILS_FILENAME = "samples_{subtask}_*.json"

TASKS = {
    "leaderboard_arc_challenge": ("ARC", "leaderboard_arc_challenge"),
    "leaderboard_bbh": ("BBH", "leaderboard_bbh"),
    "leaderboard_gpqa": ("GPQA", "leaderboard_gpqa"),
    "leaderboard_ifeval": ("IFEval", "leaderboard_ifeval"),
    "leaderboard_math_hard": ("MATH", "leaderboard_math"),
    "leaderboard_mmlu_pro": ("MMLU-Pro", "leaderboard_mmlu_pro"),
    "leaderboard_musr": ("MuSR", "leaderboard_musr"),
}
SUBTASKS = {
    "leaderboard_arc_challenge": ["leaderboard_arc_challenge"],
    "leaderboard_bbh": [
        "leaderboard_bbh_boolean_expressions",
        "leaderboard_bbh_causal_judgement",
        "leaderboard_bbh_date_understanding",
        "leaderboard_bbh_disambiguation_qa",
        "leaderboard_bbh_formal_fallacies",
        "leaderboard_bbh_geometric_shapes",
        "leaderboard_bbh_hyperbaton",
        "leaderboard_bbh_logical_deduction_five_objects",
        "leaderboard_bbh_logical_deduction_seven_objects",
        "leaderboard_bbh_logical_deduction_three_objects",
        "leaderboard_bbh_movie_recommendation",
        "leaderboard_bbh_navigate",
        "leaderboard_bbh_object_counting",
        "leaderboard_bbh_penguins_in_a_table",
        "leaderboard_bbh_reasoning_about_colored_objects",
        "leaderboard_bbh_ruin_names",
        "leaderboard_bbh_salient_translation_error_detection",
        "leaderboard_bbh_snarks", "leaderboard_bbh_sports_understanding",
        "leaderboard_bbh_temporal_sequences",
        "leaderboard_bbh_tracking_shuffled_objects_five_objects",
        "leaderboard_bbh_tracking_shuffled_objects_seven_objects",
        "leaderboard_bbh_tracking_shuffled_objects_three_objects",
        "leaderboard_bbh_web_of_lies",
    ],
    "leaderboard_gpqa": [
        "leaderboard_gpqa_extended",
        "leaderboard_gpqa_diamond",
        "leaderboard_gpqa_main",
    ],
    "leaderboard_ifeval": ["leaderboard_ifeval"],
    # "leaderboard_math_hard": [
    "leaderboard_math": [
        "leaderboard_math_algebra_hard",
        "leaderboard_math_counting_and_prob_hard",
        "leaderboard_math_geometry_hard",
        "leaderboard_math_intermediate_algebra_hard",
        "leaderboard_math_num_theory_hard",
        "leaderboard_math_prealgebra_hard",
        "leaderboard_math_precalculus_hard",
    ],
    "leaderboard_mmlu_pro": ["leaderboard_mmlu_pro"],
    "leaderboard_musr": [
        "leaderboard_musr_murder_mysteries",
        "leaderboard_musr_object_placements",
        "leaderboard_musr_team_allocation",
    ],
}


fs = HfFileSystem()


def fetch_result_paths():
    paths = fs.glob(f"{RESULTS_DATASET_ID}/**/**/*.json")
    return paths


def filter_latest_result_path_per_model(paths):
    from collections import defaultdict

    d = defaultdict(list)
    for path in paths:
        model_id, _ = path[len(RESULTS_DATASET_ID) +1:].rsplit("/", 1)
        d[model_id].append(path)
    return {model_id: max(paths) for model_id, paths in d.items()}


def get_result_path_from_model(model_id, result_path_per_model):
    return result_path_per_model[model_id]


def update_load_results_component():
    return gr.Button("Load Results", interactive=True)


def load_data(result_path) -> pd.DataFrame:
    with fs.open(result_path, "r") as f:
        data = json.load(f)
    return data


def load_results_dataframe(model_id):
    if not model_id:
        return
    result_path = get_result_path_from_model(model_id, latest_result_path_per_model)
    data = load_data(result_path)
    model_name = data.get("model_name", "Model")
    df = pd.json_normalize([{key: value for key, value in data.items() if key not in EXCLUDED_KEYS}])
    # df.columns = df.columns.str.split(".")  # .split return a list instead of a tuple
    return df.set_index(pd.Index([model_name])).reset_index()


def load_results_dataframes(*model_ids):
    return [load_results_dataframe(model_id) for model_id in model_ids]


def display_results(df_1, df_2, task):
    df = pd.concat([df.set_index("index") for df in [df_1, df_2] if "index" in df.columns])
    df = df.T.rename_axis(columns=None)
    return display_tab("results", df, task), display_tab("configs", df, task)


def display_tab(tab, df, task):
    df = df.style.format(na_rep="")
    df.hide(
        [
            row
            for row in df.index
            if (
                not row.startswith(f"{tab}.")
                or row.startswith(f"{tab}.leaderboard.")
                or row.endswith(".alias")
                or (not row.startswith(f"{tab}.{task}") if task != "All" else False)
            )
        ],
        axis="index",
    )
    start = len(f"{tab}.leaderboard_") if task == "All" else len(f"{tab}.{task} ")
    df.format_index(lambda idx: idx[start:].removesuffix(",none"), axis="index")
    return df.to_html()


def update_tasks_component():
    return gr.Radio(
            ["All"] + list(TASKS.values()),
            label="Tasks",
            info="Evaluation tasks to be displayed",
            value="All",
            interactive=True,
        )


def update_subtasks_component(task):
    return gr.Radio(
        SUBTASKS.get(task),
        info="Evaluation subtasks to be displayed",
        value=None,
    )


def update_load_details_component(model_id_1, model_id_2, subtask):
    if (model_id_1 or model_id_2) and subtask:
        return gr.Button("Load Details", interactive=True)
    else:
        return gr.Button("Load Details", interactive=False)


def load_details_dataframe(model_id, subtask):
    if not model_id or not subtask:
        return
    model_name_sanitized = model_id.replace("/", "__")
    paths = fs.glob(
        f"{DETAILS_DATASET_ID}/**/{DETAILS_FILENAME}".format(
            model_name_sanitized=model_name_sanitized, subtask=subtask
        )
    )
    if not paths:
        return
    path = max(paths)
    with fs.open(path, "r") as f:
        data = [json.loads(line) for line in f]
    df = pd.json_normalize(data)
    # df = df.rename_axis("Parameters", axis="columns")
    df["model_name"] = model_id  # Keep model_name
    return df
    # return df.set_index(pd.Index([model_id])).reset_index()


def load_details_dataframes(subtask, *model_ids):
    return [load_details_dataframe(model_id, subtask) for model_id in model_ids]


def display_details(sample_idx, *dfs):
    rows = [df.iloc[sample_idx] for df in dfs if "model_name" in df.columns and sample_idx < len(df)]
    if not rows:
        return
    # Pop model_name and add it to the column name
    df = pd.concat([row.rename(row.pop("model_name")) for row in rows], axis="columns")
    return (
        df.style
        .format(na_rep="")
        # .hide(axis="index")
        .to_html()
    )


# if __name__ == "__main__":
latest_result_path_per_model = filter_latest_result_path_per_model(fetch_result_paths())

with gr.Blocks(fill_height=True) as demo:
    gr.HTML("<h1 style='text-align: center;'>Compare Results of the 🤗 Open LLM Leaderboard</h1>")
    gr.HTML("<h3 style='text-align: center;'>Select 2 models to load and compare their results</h3>")

    with gr.Row():
        with gr.Column():
            model_id_1 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Models")
            dataframe_1 = gr.Dataframe(visible=False)
        with gr.Column():
            model_id_2 = gr.Dropdown(choices=list(latest_result_path_per_model.keys()), label="Models")
            dataframe_2 = gr.Dataframe(visible=False)

    with gr.Row():
        # with gr.Tab("All"):
        #     pass
        with gr.Tab("Results"):
            task = gr.Radio(
                ["All"] + list(TASKS.values()),
                label="Tasks",
                info="Evaluation tasks to be displayed",
                value="All",
                interactive=False,
            )
            load_results_btn = gr.Button("Load Results", interactive=False)
            with gr.Tab("Results"):
                results = gr.HTML()
            with gr.Tab("Configs"):
                configs = gr.HTML()
        with gr.Tab("Details"):
            details_task = gr.Radio(
                ["All"] + list(TASKS.values()),
                label="Tasks",
                info="Evaluation tasks to be displayed",
                value="All",
                interactive=True,
            )
            subtask = gr.Radio(
                SUBTASKS.get(details_task.value),
                label="Subtasks",
                info="Evaluation subtasks to be displayed (choose one of the Tasks above)",
            )
            sample_idx = gr.Number(value=0, label="Sample Index", info="Index of the sample to be displayed", minimum=0)
            load_details_btn = gr.Button("Load Details", interactive=False)
            details = gr.HTML()
            details_dataframe_1 = gr.Dataframe(visible=False)
            details_dataframe_2 = gr.Dataframe(visible=False)
            details_dataframe = gr.DataFrame(visible=False)

    model_id_1.change(
        fn=update_load_results_component,
        outputs=load_results_btn,
    )
    load_results_btn.click(
        fn=load_results_dataframes,
        inputs=[model_id_1, model_id_2],
        outputs=[dataframe_1, dataframe_2],
    ).then(
        fn=display_results,
        inputs=[dataframe_1, dataframe_2, task],
        outputs=[results, configs],
    ).then(
        fn=update_tasks_component,
        outputs=task,
    )
    task.change(
        fn=display_results,
        inputs=[dataframe_1, dataframe_2, task],
        outputs=[results, configs],
    )

    details_task.change(
        fn=update_subtasks_component,
        inputs=details_task,
        outputs=subtask,
    )
    gr.on(
        triggers=[model_id_1.change, model_id_2.change, subtask.change, details_task.change],
        fn=update_load_details_component,
        inputs=[model_id_1, model_id_2, subtask],
        outputs=load_details_btn,
    )
    load_details_btn.click(
        fn=load_details_dataframes,
        inputs=[subtask, model_id_1, model_id_2],
        outputs=[details_dataframe_1, details_dataframe_2],
    ).then(
        fn=display_details,
        inputs=[sample_idx, details_dataframe_1, details_dataframe_2],
        outputs=details,
    )
    sample_idx.change(
        fn=display_details,
        inputs=[sample_idx, details_dataframe_1, details_dataframe_2],
        outputs=details,
    )

demo.launch()