File size: 15,014 Bytes
9346f1c
4596a70
 
9346f1c
 
 
4596a70
 
35a0978
4596a70
 
db6f218
1f60a20
9346f1c
 
1f60a20
a460f7a
1f60a20
0a3d32f
10f9b3c
0a3d32f
10f9b3c
f742519
 
 
 
 
 
 
 
 
 
 
 
 
a885f09
f742519
9346f1c
1f60a20
9346f1c
 
 
 
 
 
a885f09
 
 
 
9346f1c
 
 
f742519
 
 
9346f1c
 
 
 
f90ad24
9346f1c
 
 
 
 
 
 
 
 
 
614ee1f
9346f1c
 
1f60a20
a885f09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
db6f218
 
 
 
1f60a20
b2c063a
a885f09
 
1363c8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a885f09
9346f1c
614ee1f
1f60a20
9346f1c
614ee1f
db6f218
614ee1f
07bfeca
 
d3fbe10
a885f09
 
 
 
 
 
 
07bfeca
 
 
d3fbe10
a885f09
 
 
 
 
 
 
07bfeca
 
614ee1f
35a0978
10f9b3c
 
 
 
 
 
 
 
 
 
35a0978
10f9b3c
1363c8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9346f1c
a885f09
1f60a20
614ee1f
1f60a20
 
a885f09
 
 
 
 
1f60a20
614ee1f
1f60a20
1363c8a
a885f09
1f60a20
 
 
614ee1f
db6f218
1f60a20
b2c063a
614ee1f
1f60a20
 
 
a885f09
 
 
 
 
1f60a20
 
 
 
614ee1f
a885f09
1f60a20
 
 
1363c8a
 
 
 
 
 
 
1f60a20
 
9346f1c
1363c8a
1f60a20
a885f09
b2c063a
1f60a20
b2c063a
1f60a20
614ee1f
1f60a20
 
 
 
614ee1f
1f60a20
85dbbc4
 
 
 
 
 
 
 
1f60a20
b2c063a
 
a095268
a885f09
85dbbc4
f742519
614ee1f
b2c063a
a885f09
85dbbc4
f742519
a885f09
1f60a20
614ee1f
1f60a20
85dbbc4
 
 
 
 
 
 
614ee1f
 
1f60a20
 
 
 
 
614ee1f
85dbbc4
1f60a20
 
614ee1f
f742519
 
 
 
1363c8a
1f60a20
 
 
614ee1f
1f60a20
 
 
 
 
 
 
a885f09
85dbbc4
f742519
1f60a20
614ee1f
1f60a20
1363c8a
 
 
 
614ee1f
10f9b3c
50a344f
 
 
 
 
1f60a20
01233b7
 
1f60a20
a885f09
 
07bfeca
e4b25b8
a460f7a
db6f218
 
 
 
614ee1f
a460f7a
db6f218
a885f09
 
614ee1f
50a344f
10f9b3c
 
9346f1c
a885f09
 
 
614ee1f
9346f1c
a885f09
 
07bfeca
614ee1f
a885f09
 
35a0978
1363c8a
 
 
 
 
 
 
35a0978
db6f218
1363c8a
 
 
 
 
 
 
35a0978
1363c8a
 
 
 
 
 
a885f09
614ee1f
1f60a20
 
a885f09
1363c8a
 
 
 
 
 
 
 
a885f09
614ee1f
1f60a20
 
b2c063a
 
 
614ee1f
b2c063a
a885f09
 
 
db6f218
b2c063a
db6f218
614ee1f
1f60a20
 
35a0978
85dbbc4
 
 
a885f09
 
 
 
 
 
 
 
 
 
 
10f9b3c
 
 
 
 
 
 
 
 
 
 
 
 
9567fa6
10f9b3c
01233b7
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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
import os
import json

import numpy as np
import gradio as gr
import pandas as pd

from apscheduler.schedulers.background import BackgroundScheduler
from content import CHANGELOG_TEXT
from huggingface_hub import Repository, HfApi
from transformers import AutoConfig
from utils import get_eval_results_dicts, make_clickable_model

# 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", None))

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


def get_all_requested_models(requested_models_dir):
    depth = 1
    file_names = []

    for root, dirs, files in os.walk(requested_models_dir):
        current_depth = root.count(os.sep) - requested_models_dir.count(os.sep)
        if current_depth == depth:
            file_names.extend([os.path.join(root, file) for file in files])

    return set([file_name.lower().split("./evals/")[1] for file_name in file_names])


repo = None
requested_models = None
if H4_TOKEN:
    print("pulling repo")
    # try:
    #     shutil.rmtree("./evals/")
    # except:
    #     pass

    repo = Repository(
        local_dir="./evals/",
        clone_from=LMEH_REPO,
        use_auth_token=H4_TOKEN,
        repo_type="dataset",
    )
    repo.git_pull()

    requested_models_dir = "./evals/eval_requests"
    requested_models = get_all_requested_models(requested_models_dir)


# parse the results
BENCHMARKS = ["arc_challenge", "hellaswag", "hendrycks", "truthfulqa_mc"]

METRICS = ["acc_norm", "acc_norm", "acc_norm", "mc2"]


def load_results(model, benchmark, metric):
    file_path = os.path.join("evals", model, f"{model}-eval_{benchmark}.json")
    if not os.path.exists(file_path):
        return 0.0, None

    with open(file_path) as fp:
        data = json.load(fp)
    accs = np.array([v[metric] for k, v in data["results"].items()])
    mean_acc = np.mean(accs)
    return mean_acc, data["config"]["model_args"]


COLS = [
    "Model",
    "Revision",
    "Average ⬆️",
    "ARC (25-shot) ⬆️",
    "HellaSwag (10-shot) ⬆️",
    "MMLU (5-shot) ⬆️",
    "TruthfulQA (0-shot) ⬆️",
]
TYPES = [
    "markdown",
    "str",
    "number",
    "number",
    "number",
    "number",
    "number",
]

if not IS_PUBLIC:
    COLS.insert(2, "8bit")
    TYPES.insert(2, "bool")

EVAL_COLS = ["model", "revision", "private", "8bit_eval", "is_delta_weight", "status"]
EVAL_TYPES = ["markdown", "str", "bool", "bool", "bool", "str"]

BENCHMARK_COLS = [
    "ARC (25-shot) ⬆️",
    "HellaSwag (10-shot) ⬆️",
    "MMLU (5-shot) ⬆️",
    "TruthfulQA (0-shot) ⬆️",
]


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():
    if repo:
        print("pulling changes")
        repo.git_pull()

    all_data = get_eval_results_dicts(IS_PUBLIC)

    if not IS_PUBLIC:
        gpt4_values = {
            "Model": f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">gpt4</a>',
            "Revision": "tech report",
            "8bit": None,
            "Average ⬆️": 84.3,
            "ARC (25-shot) ⬆️": 96.3,
            "HellaSwag (10-shot) ⬆️": 95.3,
            "MMLU (5-shot) ⬆️": 86.4,
            "TruthfulQA (0-shot) ⬆️": 59.0,
        }
        all_data.append(gpt4_values)
        gpt35_values = {
            "Model": f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">gpt3.5</a>',
            "Revision": "tech report",
            "8bit": None,
            "Average ⬆️": 71.9,
            "ARC (25-shot) ⬆️": 85.2,
            "HellaSwag (10-shot) ⬆️": 85.5,
            "MMLU (5-shot) ⬆️": 70.0,
            "TruthfulQA (0-shot) ⬆️": 47.0,
        }
        all_data.append(gpt35_values)

    base_line = {
        "Model": "<p>Baseline</p>",
        "Revision": "N/A",
        "8bit": None,
        "Average ⬆️": 25.0,
        "ARC (25-shot) ⬆️": 25.0,
        "HellaSwag (10-shot) ⬆️": 25.0,
        "MMLU (5-shot) ⬆️": 25.0,
        "TruthfulQA (0-shot) ⬆️": 25.0,
    }

    all_data.append(base_line)

    df = pd.DataFrame.from_records(all_data)
    df = df.sort_values(by=["Average ⬆️"], ascending=False)
    df = df[COLS]

    # get incomplete models
    incomplete_models = df[has_nan_values(df, BENCHMARK_COLS)]["Model"].tolist()
    print(
        [
            model.split(" style")[0].split("https://huggingface.co/")[1]
            for model in incomplete_models
        ]
    )

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


def get_eval_table():
    if repo:
        print("pulling changes for eval")
        repo.git_pull()
    entries = [
        entry
        for entry in os.listdir("evals/eval_requests")
        if not entry.startswith(".")
    ]
    all_evals = []

    for entry in entries:
        # print(entry)
        if ".json" in entry:
            file_path = os.path.join("evals/eval_requests", 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"evals/eval_requests/{entry}")
                if not e.startswith(".")
            ]
            for sub_entry in sub_entries:
                file_path = os.path.join("evals/eval_requests", 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]


leaderboard = get_leaderboard()
finished_eval_queue, running_eval_queue, pending_eval_queue = get_eval_table()


def is_model_on_hub(model_name, revision) -> bool:
    try:
        config = AutoConfig.from_pretrained(model_name, revision=revision)
        return True

    except Exception as e:
        print("Could not get the model config from the hub")
        print(e)
        return False


def add_new_eval(
    model: str,
    base_model: str,
    revision: str,
    is_8_bit_eval: bool,
    private: bool,
    is_delta_weight: bool,
):
    # check the model actually exists before adding the eval
    if revision == "":
        revision = "main"
    if is_delta_weight and not is_model_on_hub(base_model, revision):
        error_message = f'Base model "{base_model}" was not found on hub!'
        print(error_message)
        return f"<p style='color: red; font-size: 20px; text-align: center;'>{error_message}</p>"

    if not is_model_on_hub(model, revision):
        error_message = f'Model "{model}"was not found on hub!'
        print(error_message)
        return f"<p style='color: red; font-size: 20px; text-align: center;'>{error_message}</p>"

    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",
    }

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

    OUT_DIR = f"eval_requests/{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.lower() in requested_models:
        duplicate_request_message = "This model has been already submitted."
        return f"<p style='color: orange; font-size: 20px; text-align: center;'>{duplicate_request_message}</p>"

    with open(out_path, "w") as f:
        f.write(json.dumps(eval_entry))
    LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"

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

    success_message = "Your request has been submitted to the evaluation queue!"
    return f"<p style='color: green; font-size: 20px; text-align: center;'>{success_message}</p>"


def refresh():
    leaderboard = get_leaderboard()
    finished_eval_queue, running_eval_queue, pending_eval_queue = get_eval_table()
    get_leaderboard(), get_eval_table()
    return leaderboard, finished_eval_queue, running_eval_queue, pending_eval_queue


custom_css = """
#changelog-text {
    font-size: 18px !important;
}
"""

demo = gr.Blocks(css=custom_css)
with demo:
    with gr.Row():
        gr.Markdown(
            f"""
# 🤗 Open LLM Leaderboard
<font size="4">With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art. The 🤗 Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. We evaluate models on 4 key benchmarks from the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank">  Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks. A key advantage of this leaderboard is that anyone from the community can submit a model for automated evaluation on the 🤗 GPU cluster, as long as it is a 🤗 Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as LLaMa.

Evaluation is performed against 4 popular benchmarks:
- <a href="https://arxiv.org/abs/1803.05457" target="_blank">  AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions.
- <a href="https://arxiv.org/abs/1905.07830" target="_blank">  HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
- <a href="https://arxiv.org/abs/2009.03300" target="_blank">  MMLU </a>  (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
- <a href="https://arxiv.org/abs/2109.07958" target="_blank">  TruthfulQA </a> (0-shot) - a benchmark to measure whether a language model is truthful in generating answers to questions.

We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings. </font>
        """
        )

    with gr.Accordion("CHANGELOG", open=False):
        changelog = gr.Markdown(CHANGELOG_TEXT, elem_id="changelog-text")

    with gr.Row():
        leaderboard_table = gr.components.Dataframe(
            value=leaderboard, headers=COLS, datatype=TYPES, max_rows=5
        )

    with gr.Row():
        gr.Markdown(
            f"""
    # Evaluation Queue for the 🤗 Open LLM Leaderboard, these models will be automatically evaluated on the 🤗 cluster

    """
        )
    with gr.Accordion("✅ Finished Evaluations", open=False):
        with gr.Row():
            finished_eval_table = gr.components.Dataframe(
                value=finished_eval_queue,
                headers=EVAL_COLS,
                datatype=EVAL_TYPES,
                max_rows=5,
            )
    with gr.Accordion("🔄 Running Evaluation Queue", open=False):
        with gr.Row():
            running_eval_table = gr.components.Dataframe(
                value=running_eval_queue,
                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,
                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)")

        with gr.Row():
            submit_button = gr.Button("Submit Eval")

        with gr.Row():
            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,
            )

    # demo.load(
    #     refresh,
    #     inputs=[],
    #     outputs=[
    #         leaderboard_table,
    #         finished_eval_table,
    #         running_eval_table,
    #         pending_eval_table,
    #     ],
    # )
    
scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, 'interval', seconds=1200)
scheduler.start()
demo.launch()