File size: 7,395 Bytes
9346f1c
 
 
 
1f60a20
 
9346f1c
 
 
1f60a20
b2c063a
1f60a20
9346f1c
 
1f60a20
 
9346f1c
 
1f60a20
9346f1c
 
 
 
 
 
1f60a20
9346f1c
 
 
 
 
 
 
f90ad24
9346f1c
 
 
 
 
 
 
 
 
 
 
 
 
1f60a20
b2c063a
 
1f60a20
b2c063a
 
9346f1c
 
1f60a20
9346f1c
 
f90ad24
9346f1c
1f60a20
9346f1c
 
 
 
1f60a20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b2c063a
1f60a20
 
 
 
 
 
 
 
 
 
 
b2c063a
1f60a20
 
 
 
 
 
 
 
9346f1c
1f60a20
 
b2c063a
1f60a20
b2c063a
1f60a20
 
 
 
 
 
 
 
 
b2c063a
1f60a20
b2c063a
 
 
 
1f60a20
 
 
 
 
 
b2c063a
1f60a20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9346f1c
 
 
1f60a20
 
 
 
 
9346f1c
 
 
 
1f60a20
 
 
9346f1c
1f60a20
 
9346f1c
1f60a20
9346f1c
1f60a20
 
 
 
 
 
 
 
 
 
 
 
b2c063a
 
 
 
 
 
 
1f60a20
 
 
b2c063a
1f60a20
 
 
9346f1c
 
1f60a20
 
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
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, get_n_params

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

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 = ["eval_name",  "total ⬆️", "ARC (25-shot) ⬆️", "HellaSwag (10-shot) ⬆️", "MMLU (5-shot) ⬆️", "TruthQA (0-shot) ⬆️", "base_model"]
TYPES = ["str",  "number", "number", "number", "number", "number","markdown", ]

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()
    dataframe = pd.DataFrame.from_records(all_data)
    dataframe = dataframe.sort_values(by=['total ⬆️'], ascending=False)
    
    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"] = get_n_params(data["model"])
            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, revision:str, private:bool, is_8_bit_eval: bool, is_delta_weight:bool):
    # check the model actually exists before adding the eval
    if revision == "":
        revision = "main"
    print("revision", revision)
    if not is_model_on_hub(model, revision):
        print(model, "not found on hub")
        return
    print("adding new eval")
    
    eval_entry = {
        "model" : 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"""
        # 🤗 H4 Model Evaluation leaderboard using the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank"> LMEH benchmark suite </a>. 
        Evaluation is performed against 4 popular benchmarks AI2 Reasoning Challenge, HellaSwag, MMLU, and TruthFul QC MC. To run your own benchmarks, refer to the README in the H4 repo.
        """)
    
    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.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")
                private = gr.Checkbox(False, label="Private")
                is_delta_weight = gr.Checkbox(False, label="Delta weights")
            
        with gr.Row():
            submit_button = gr.Button("Submit Eval")
            submit_button.click(add_new_eval, [model_name_textbox, revision_name_textbox, is_8bit_toggle, private, is_delta_weight])
        
        
        
    


print("adding refresh leaderboard")
def refresh_leaderboard():
    leaderboard_table = get_leaderboard()
    print("leaderboard updated")

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

block.launch()