File size: 10,522 Bytes
9346f1c
 
 
 
1f60a20
 
9346f1c
 
 
1f60a20
db6f218
1f60a20
9346f1c
 
1f60a20
db6f218
1f60a20
9346f1c
 
1f60a20
9346f1c
 
 
 
 
 
1f60a20
9346f1c
 
 
 
 
 
 
f90ad24
9346f1c
 
 
 
 
 
 
 
 
 
 
 
 
1f60a20
db6f218
 
 
 
 
 
1f60a20
b2c063a
 
9346f1c
 
1f60a20
9346f1c
 
db6f218
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9346f1c
1f60a20
fcb01e3
9346f1c
 
 
1f60a20
 
 
 
 
 
 
 
 
 
 
 
 
 
db6f218
1f60a20
b2c063a
1f60a20
 
 
 
 
 
 
 
 
 
 
b2c063a
1f60a20
 
 
 
 
 
 
 
9346f1c
1f60a20
 
b2c063a
1f60a20
b2c063a
1f60a20
 
 
 
 
 
 
 
 
5cb1426
1f60a20
b2c063a
 
a095268
 
 
 
b2c063a
1f60a20
 
 
 
 
 
a095268
b2c063a
1f60a20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9346f1c
 
 
1f60a20
 
db6f218
 
 
 
 
 
 
 
 
 
1f60a20
9346f1c
 
 
 
1f60a20
 
 
9346f1c
1f60a20
 
9346f1c
1f60a20
db6f218
 
 
 
1f60a20
 
 
 
 
 
 
 
 
b2c063a
 
 
db6f218
b2c063a
db6f218
 
b2c063a
db6f218
1f60a20
 
 
a095268
1f60a20
 
 
9346f1c
 
1f60a20
 
9346f1c
 
db6f218
 
9346f1c
 
 
1f60a20
 
 
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
import os
import shutil
import numpy as np
import gradio as gr
from huggingface_hub import Repository, HfApi
from transformers import AutoConfig
import json
from apscheduler.schedulers.background import BackgroundScheduler
import pandas as pd
import datetime
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(True) # add secret here

repo=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()


# 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 = ["base_model", "revision", "total ⬆️", "ARC (25-shot) ⬆️", "HellaSwag (10-shot) ⬆️", "MMLU (5-shot) ⬆️", "TruthQA (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"]
def get_leaderboard():
    if repo: 
        print("pulling changes")
        repo.git_pull()
        
    all_data = get_eval_results_dicts(IS_PUBLIC)
    
    gpt4_values = {
        "base_model":f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt4</a>', 
        "revision":"tech report", 
        "8bit":None,
        "total ⬆️":84.3,
        "ARC (25-shot) ⬆️":96.3,
        "HellaSwag (10-shot) ⬆️":95.3,
        "MMLU (5-shot) ⬆️":86.4,
        "TruthQA (0-shot) ⬆️":59.0,
    }
    all_data.append(gpt4_values)
    gpt35_values = {
        "base_model":f'<a target="_blank" href=https://arxiv.org/abs/2303.08774 style="color: blue; text-decoration: underline;text-decoration-style: dotted;">gpt3.5</a>', 
        "revision":"tech report", 
        "8bit":None,
        "total ⬆️":71.9,
        "ARC (25-shot) ⬆️":85.2,
        "HellaSwag (10-shot) ⬆️":85.5,
        "MMLU (5-shot) ⬆️":70.0,
        "TruthQA (0-shot) ⬆️":47.0,
    }
    all_data.append(gpt35_values)
    
    dataframe = pd.DataFrame.from_records(all_data)
    dataframe = dataframe.sort_values(by=['total ⬆️'], ascending=False)
    print(dataframe)
    dataframe = dataframe[COLS]
    return dataframe

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)

    
    dataframe = pd.DataFrame.from_records(all_evals)
    return dataframe[EVAL_COLS]


leaderboard = get_leaderboard()
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):
        print(base_model, "base model not found on hub")
        return
    
    if not is_model_on_hub(model, revision):
        print(model, "not found on hub")
        return
    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"
    
    with open(out_path, "w") as f:
        f.write(json.dumps(eval_entry))
    LMEH_REPO = "HuggingFaceH4/lmeh_evaluations"
    
    api = HfApi()
    api.upload_file(
        path_or_fileobj=out_path,
        path_in_repo=out_path,
        repo_id=LMEH_REPO,
        token=H4_TOKEN,
        repo_type="dataset",
    )

    
def refresh():
    return get_leaderboard(), get_eval_table()
    


block = gr.Blocks()
with block: 
    with gr.Row():
        gr.Markdown(f"""
# 🤗 Open Chatbot Leaderboard
<font size="4">With the plethora of chatbot LLMs 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 chatbot is the current state of the art. The 🤗 Open Chatbot Leaderboard aims to track, rank and evaluate chatbot models as they are released. We evaluate models of 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 🤗 research cluster. As long as it is Transformers model with weights on the 🤗 hub. We also support delta-weights for non-commercial licensed models, such as llama.
<p>
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">  Truthful QA MC </a> (0-shot) - a benchmark to measure whether a language model is truthful in generating answers to questions. 
<p>
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.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 LMEH benchmarks, these models will be automatically evaluated on the 🤗 cluster
    
    """)
    with gr.Accordion("Evaluation Queue", open=False):
        with gr.Row():
            eval_table = gr.components.Dataframe(value=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, eval_table]) 
        
    with gr.Accordion("Submit a new model for evaluation"):
        # with gr.Row():
        #     gr.Markdown(f"""# 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")
            submit_button.click(add_new_eval, [model_name_textbox, base_model_name_textbox, revision_name_textbox, is_8bit_toggle, private, is_delta_weight])
        
        
        
    


print("adding refresh leaderboard")
def refresh_leaderboard():
    leaderboard_table = get_leaderboard()
    eval_table = get_eval_table()
    print("refreshing leaderboard")

scheduler = BackgroundScheduler()
scheduler.add_job(func=refresh_leaderboard, trigger="interval", seconds=300) # refresh every 5 mins
scheduler.start()

block.launch()