File size: 13,924 Bytes
4596a70
0227006
d35aee2
4596a70
0227006
9346f1c
0227006
9346f1c
4596a70
460d762
4596a70
0227006
460d762
 
 
 
 
 
 
1f60a20
9346f1c
 
1f60a20
460d762
 
1f60a20
d52179b
 
0a3d32f
2a73469
 
10f9b3c
8cb7546
 
 
10f9b3c
6e8f400
f742519
d52179b
 
 
 
f742519
460d762
 
 
9346f1c
460d762
 
9346f1c
460d762
1f60a20
1363c8a
 
 
 
 
 
 
 
2a73469
ffefe11
460d762
2a73469
460d762
614ee1f
db6f218
614ee1f
07bfeca
 
 
614ee1f
460d762
 
10f9b3c
1363c8a
460d762
1363c8a
 
 
 
 
9346f1c
a885f09
ffefe11
460d762
 
2a73469
460d762
2a73469
a885f09
 
d52179b
a885f09
 
1f60a20
614ee1f
1f60a20
a885f09
d52179b
1f60a20
 
614ee1f
db6f218
1f60a20
b2c063a
614ee1f
1f60a20
 
 
a885f09
 
d52179b
a885f09
 
1f60a20
d52179b
1f60a20
 
614ee1f
a885f09
1f60a20
 
 
1363c8a
 
 
 
 
 
 
1f60a20
 
0227006
ffefe11
 
 
 
 
 
 
a885f09
b2c063a
1f60a20
460d762
 
 
 
 
614ee1f
1f60a20
460d762
 
614ee1f
1f60a20
85dbbc4
 
 
 
 
 
 
 
8696209
 
1f60a20
b2c063a
 
614ee1f
460d762
 
 
 
 
 
 
 
a885f09
1f60a20
614ee1f
1f60a20
85dbbc4
 
 
 
 
 
 
8696209
614ee1f
 
1f60a20
 
 
 
 
614ee1f
d52179b
1f60a20
 
614ee1f
f742519
460d762
 
1363c8a
1f60a20
 
614ee1f
1f60a20
 
 
 
 
 
 
a885f09
460d762
1f60a20
614ee1f
1f60a20
ffefe11
 
 
 
 
 
 
 
 
 
 
 
614ee1f
2a73469
aa7c3f4
460d762
aa7c3f4
 
7644705
8cb7546
7644705
 
 
 
 
 
 
 
8cb7546
6a6e05c
 
aa7c3f4
7644705
01233b7
 
58733e4
6e8f400
0227006
6e8f400
 
 
 
 
 
10f9b3c
8cb7546
6e8f400
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0227006
6e8f400
 
 
 
 
 
 
 
460d762
6e8f400
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d2e8eca
 
 
 
 
 
 
0227006
6e8f400
 
 
 
 
 
0227006
6e8f400
 
 
 
 
0227006
6e8f400
 
 
 
 
 
 
0227006
6e8f400
 
 
 
 
 
 
 
0227006
 
6e8f400
 
 
 
 
 
 
 
 
 
 
0227006
6e8f400
 
 
 
 
 
 
0227006
6e8f400
 
 
 
 
 
 
 
 
 
 
8cb7546
6e8f400
 
 
 
 
 
 
 
 
 
 
 
 
8cb7546
 
 
 
 
 
 
 
 
10f9b3c
 
c131125
10f9b3c
f458f0b
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
import json
import os
from datetime import datetime, timezone


import gradio as gr
import numpy as np
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import HfApi
from transformers import AutoConfig

from src.auto_leaderboard.get_model_metadata import apply_metadata
from src.assets.text_content import *
from src.auto_leaderboard.load_results import get_eval_results_dicts, make_clickable_model
from src.assets.hardcoded_evals import gpt4_values, gpt35_values, baseline
from src.assets.css_html_js import custom_css, get_window_url_params
from src.utils_display import AutoEvalColumn, EvalQueueColumn, EloEvalColumn, fields, styled_error, styled_warning, styled_message
from src.init import load_all_info_from_hub

# clone / pull the lmeh eval data
H4_TOKEN = os.environ.get("H4_TOKEN", None)
LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
IS_PUBLIC = bool(os.environ.get("IS_PUBLIC", True))
ADD_PLOTS = False

EVAL_REQUESTS_PATH = "auto_evals/eval_requests"

api = HfApi()


def restart_space():
    api.restart_space(
        repo_id="HuggingFaceH4/open_llm_leaderboard", token=H4_TOKEN
    )

auto_eval_repo, requested_models = load_all_info_from_hub(LMEH_REPO)

COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]
TYPES = [c.type for c in fields(AutoEvalColumn) if not c.hidden]
COLS_LITE = [c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]
TYPES_LITE = [c.type for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden]

if not IS_PUBLIC:
    COLS.insert(2, AutoEvalColumn.is_8bit.name)
    TYPES.insert(2, AutoEvalColumn.is_8bit.type)

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [c.name for c in [AutoEvalColumn.arc, AutoEvalColumn.hellaswag, AutoEvalColumn.mmlu, AutoEvalColumn.truthfulqa]]


def has_no_nan_values(df, columns):
    return df[columns].notna().all(axis=1)


def has_nan_values(df, columns):
    return df[columns].isna().any(axis=1)


def get_leaderboard_df():
    if auto_eval_repo:
        print("Pulling evaluation results for the leaderboard.")
        auto_eval_repo.git_pull()

    all_data = get_eval_results_dicts(IS_PUBLIC)

    if not IS_PUBLIC:
        all_data.append(gpt4_values)
        all_data.append(gpt35_values)

    all_data.append(baseline)
    apply_metadata(all_data)  # Populate model type based on known hardcoded values in `metadata.py`

    df = pd.DataFrame.from_records(all_data)
    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
    df = df[COLS]

    # filter out if any of the benchmarks have not been produced
    df = df[has_no_nan_values(df, BENCHMARK_COLS)]
    return df


def get_evaluation_queue_df():
    # todo @saylortwift: replace the repo by the one you created for the eval queue
    if auto_eval_repo:
        print("Pulling changes for the evaluation queue.")
        auto_eval_repo.git_pull()

    entries = [
        entry
        for entry in os.listdir(EVAL_REQUESTS_PATH)
        if not entry.startswith(".")
    ]
    all_evals = []

    for entry in entries:
        if ".json" in entry:
            file_path = os.path.join(EVAL_REQUESTS_PATH, entry)
            with open(file_path) as fp:
                data = json.load(fp)

            data["# params"] = "unknown"
            data["model"] = make_clickable_model(data["model"])
            data["revision"] = data.get("revision", "main")

            all_evals.append(data)
        else:
            # this is a folder
            sub_entries = [
                e
                for e in os.listdir(f"{EVAL_REQUESTS_PATH}/{entry}")
                if not e.startswith(".")
            ]
            for sub_entry in sub_entries:
                file_path = os.path.join(EVAL_REQUESTS_PATH, entry, sub_entry)
                with open(file_path) as fp:
                    data = json.load(fp)

                # data["# params"] = get_n_params(data["model"])
                data["model"] = make_clickable_model(data["model"])
                all_evals.append(data)

    pending_list = [e for e in all_evals if e["status"] == "PENDING"]
    running_list = [e for e in all_evals if e["status"] == "RUNNING"]
    finished_list = [e for e in all_evals if e["status"] == "FINISHED"]
    df_pending = pd.DataFrame.from_records(pending_list)
    df_running = pd.DataFrame.from_records(running_list)
    df_finished = pd.DataFrame.from_records(finished_list)
    return df_finished[EVAL_COLS], df_running[EVAL_COLS], df_pending[EVAL_COLS]



original_df = get_leaderboard_df()
leaderboard_df = original_df.copy()
(
    finished_eval_queue_df,
    running_eval_queue_df,
    pending_eval_queue_df,
) = get_evaluation_queue_df()

def is_model_on_hub(model_name, revision) -> bool:
    try:
        AutoConfig.from_pretrained(model_name, revision=revision)
        return True, None
    
    except ValueError as e:
        return False, "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."

    except Exception as e:
        print("Could not get the model config from the hub.: \n", e)
        return False, "was not found on hub!"


def add_new_eval(
    model: str,
    base_model: str,
    revision: str,
    is_8_bit_eval: bool,
    private: bool,
    is_delta_weight: bool,
):
    current_time = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")

    # check the model actually exists before adding the eval
    if revision == "":
        revision = "main"

    if is_delta_weight: 
        base_model_on_hub, error = is_model_on_hub(base_model, revision)
        if not base_model_on_hub:
            return styled_error(f'Base model "{base_model}" {error}')

    model_on_hub, error = is_model_on_hub(model, revision)
    if not model_on_hub:
        return styled_error(f'Model "{model}" {error}')

    print("adding new eval")

    eval_entry = {
        "model": model,
        "base_model": base_model,
        "revision": revision,
        "private": private,
        "8bit_eval": is_8_bit_eval,
        "is_delta_weight": is_delta_weight,
        "status": "PENDING",
        "submitted_time": current_time,
    }

    user_name = ""
    model_path = model
    if "/" in model:
        user_name = model.split("/")[0]
        model_path = model.split("/")[1]

    OUT_DIR = f"{EVAL_REQUESTS_PATH}/{user_name}"
    os.makedirs(OUT_DIR, exist_ok=True)
    out_path = f"{OUT_DIR}/{model_path}_eval_request_{private}_{is_8_bit_eval}_{is_delta_weight}.json"

    # Check for duplicate submission
    if out_path.split("eval_requests/")[1].lower() in requested_models:
        return styled_warning("This model has been already submitted.")

    with open(out_path, "w") as f:
        f.write(json.dumps(eval_entry))

    api.upload_file(
        path_or_fileobj=out_path,
        path_in_repo=out_path,
        repo_id=LMEH_REPO,
        token=H4_TOKEN,
        repo_type="dataset",
    )

    return styled_message("Your request has been submitted to the evaluation queue!")


def refresh():
    leaderboard_df = get_leaderboard_df()
    (
        finished_eval_queue_df,
        running_eval_queue_df,
        pending_eval_queue_df,
    ) = get_evaluation_queue_df()
    return (
        leaderboard_df,
        finished_eval_queue_df,
        running_eval_queue_df,
        pending_eval_queue_df,
    )


def search_table(df, query):
    filtered_df = df[df[AutoEvalColumn.dummy.name].str.contains(query, case=False)]
    return filtered_df


def change_tab(query_param):
    query_param = query_param.replace("'", '"')
    query_param = json.loads(query_param)

    if (
        isinstance(query_param, dict)
        and "tab" in query_param
        and query_param["tab"] == "evaluation"
    ):
        return gr.Tabs.update(selected=1)
    else:
        return gr.Tabs.update(selected=0)


demo = gr.Blocks(css=custom_css)
with demo:
    gr.HTML(TITLE)
    gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
    with gr.Row():
        with gr.Box(elem_id="search-bar-table-box"):
            search_bar = gr.Textbox(
                placeholder="πŸ” Search your model and press ENTER...",
                show_label=False,
                elem_id="search-bar",
            )

    with gr.Tabs(elem_classes="tab-buttons") as tabs:
        with gr.TabItem("πŸ… LLM Benchmark (lite)", elem_id="llm-benchmark-tab-table", id=0):
            leaderboard_table_lite = gr.components.Dataframe(
                value=leaderboard_df[COLS_LITE],
                headers=COLS_LITE,
                datatype=TYPES_LITE,
                max_rows=None,
                elem_id="leaderboard-table-lite",
            )
            # Dummy leaderboard for handling the case when the user uses backspace key
            hidden_leaderboard_table_for_search_lite = gr.components.Dataframe(
                value=original_df[COLS_LITE],
                headers=COLS_LITE,
                datatype=TYPES_LITE,
                max_rows=None,
                visible=False,
            )
            search_bar.submit(
                search_table,
                [hidden_leaderboard_table_for_search_lite, search_bar],
                leaderboard_table_lite,
            )

        with gr.TabItem("πŸ“Š Extended view", elem_id="llm-benchmark-tab-table", id=1):
            leaderboard_table = gr.components.Dataframe(
                value=leaderboard_df,
                headers=COLS,
                datatype=TYPES,
                max_rows=None,
                elem_id="leaderboard-table",
            )

            # Dummy leaderboard for handling the case when the user uses backspace key
            hidden_leaderboard_table_for_search = gr.components.Dataframe(
                value=original_df,
                headers=COLS,
                datatype=TYPES,
                max_rows=None,
                visible=False,
            )
            search_bar.submit(
                search_table,
                [hidden_leaderboard_table_for_search, search_bar],
                leaderboard_table,
            )
        with gr.TabItem("About", elem_id="llm-benchmark-tab-table", id=2):
            gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
            with gr.Accordion("πŸ“™ Citation", open=False):
                citation_button = gr.Textbox(
                    value=CITATION_BUTTON_TEXT,
                    label=CITATION_BUTTON_LABEL,
                    elem_id="citation-button",
                ).style(show_copy_button=True)


    with gr.Column():
        with gr.Row():
            gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")

        with gr.Column():
            with gr.Accordion("βœ… Finished Evaluations", open=False):
                with gr.Row():
                    finished_eval_table = gr.components.Dataframe(
                        value=finished_eval_queue_df,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        max_rows=5,
                    )
            with gr.Accordion("πŸ”„ Running Evaluation Queue", open=False):
                with gr.Row():
                    running_eval_table = gr.components.Dataframe(
                        value=running_eval_queue_df,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        max_rows=5,
                    )

            with gr.Accordion("⏳ Pending Evaluation Queue", open=False):
                with gr.Row():
                    pending_eval_table = gr.components.Dataframe(
                        value=pending_eval_queue_df,
                        headers=EVAL_COLS,
                        datatype=EVAL_TYPES,
                        max_rows=5,
                    )

        with gr.Row():
            refresh_button = gr.Button("Refresh")
            refresh_button.click(
                refresh,
                inputs=[],
                outputs=[
                    leaderboard_table,
                    finished_eval_table,
                    running_eval_table,
                    pending_eval_table,
                ],
            )
        with gr.Accordion("Submit a new model for evaluation"):
            with gr.Row():
                with gr.Column():
                    model_name_textbox = gr.Textbox(label="Model name")
                    revision_name_textbox = gr.Textbox(
                        label="revision", placeholder="main"
                    )

                with gr.Column():
                    is_8bit_toggle = gr.Checkbox(
                        False, label="8 bit eval", visible=not IS_PUBLIC
                    )
                    private = gr.Checkbox(
                        False, label="Private", visible=not IS_PUBLIC
                    )
                    is_delta_weight = gr.Checkbox(False, label="Delta weights")
                    base_model_name_textbox = gr.Textbox(
                        label="base model (for delta)"
                    )

            submit_button = gr.Button("Submit Eval")
            submission_result = gr.Markdown()
            submit_button.click(
                add_new_eval,
                [
                    model_name_textbox,
                    base_model_name_textbox,
                    revision_name_textbox,
                    is_8bit_toggle,
                    private,
                    is_delta_weight,
                ],
                submission_result,
            )

    dummy = gr.Textbox(visible=False)
    demo.load(
        change_tab,
        dummy,
        tabs,
        _js=get_window_url_params,
    )

scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "interval", seconds=3600)
scheduler.start()
demo.queue(concurrency_count=40).launch()