File size: 10,227 Bytes
9d22eee
2a5f9fb
6b87e28
a5d34d3
 
df66f6e
 
0a3530a
a5d34d3
092c345
 
a5d34d3
 
 
 
 
 
 
092c345
a5d34d3
 
 
 
 
 
 
 
 
092c345
6b87e28
 
 
 
 
 
 
 
 
 
9d22eee
 
 
 
 
 
 
 
 
 
0a3530a
9d22eee
beb2b32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
0a3530a
2a5f9fb
 
 
b202e95
2a5f9fb
 
 
 
 
 
 
 
0a3530a
9d22eee
 
 
 
0a3530a
9d22eee
 
beb2b32
 
9d22eee
 
 
 
 
b762711
9d22eee
 
 
0a3530a
 
 
9d22eee
bcd77eb
9b2e755
b4ba8b7
beb2b32
 
 
 
b7d036c
9d22eee
 
 
2a5f9fb
0a3530a
2a5f9fb
 
beb2b32
 
2a5f9fb
beb2b32
2a5f9fb
beb2b32
2a5f9fb
 
 
beb2b32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
b1a1395
 
 
 
 
 
 
 
beb2b32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
0a3530a
2a5f9fb
9d22eee
2a5f9fb
0a3530a
2a5f9fb
 
 
10a491c
 
 
 
 
beb2b32
2a5f9fb
 
 
 
 
beb2b32
 
2a5f9fb
beb2b32
9d6aecc
beb2b32
2a5f9fb
beb2b32
05bda40
beb2b32
05bda40
2a5f9fb
 
0a3530a
9d22eee
 
 
 
 
0a3530a
9d22eee
 
 
 
 
 
 
 
beb2b32
9d22eee
 
 
 
 
 
 
 
 
 
 
 
2a5f9fb
 
 
9b2e755
 
2a5f9fb
 
 
 
b1a1395
2a5f9fb
 
 
 
 
 
 
 
 
 
 
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
from dataclasses import dataclass, make_dataclass
from enum import Enum
import json
import logging
from datetime import datetime
import pandas as pd


# Configure logging
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")


def parse_datetime(datetime_str):
    formats = [
        "%Y-%m-%dT%H-%M-%S.%f",  # Format with dashes
        "%Y-%m-%dT%H:%M:%S.%f",  # Standard format with colons
        "%Y-%m-%dT%H %M %S.%f",  # Spaces as separator
    ]

    for fmt in formats:
        try:
            return datetime.strptime(datetime_str, fmt)
        except ValueError:
            continue
    # in rare cases set unix start time for files with incorrect time (legacy files)
    logging.error(f"No valid date format found for: {datetime_str}")
    return datetime(1970, 1, 1)


def load_json_data(file_path):
    """Safely load JSON data from a file."""
    try:
        with open(file_path, "r") as file:
            return json.load(file)
    except json.JSONDecodeError:
        print(f"Error reading JSON from {file_path}")
        return None  # Or raise an exception


def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str


class Tasks(Enum):
    ifeval = Task("leaderboard_ifeval", "strict_acc,none", "IFEval")
    ifeval_raw = Task("leaderboard_ifeval", "strict_acc,none", "IFEval Raw")

    bbh = Task("leaderboard_bbh", "acc_norm,none", "BBH")
    bbh_raw = Task("leaderboard_bbh", "acc_norm,none", "BBH Raw")

    math = Task("leaderboard_math_hard", "exact_match,none", "MATH Lvl 5")
    math_raw = Task("leaderboard_math_hard", "exact_match,none", "MATH Lvl 5 Raw")

    gpqa = Task("leaderboard_gpqa", "acc_norm,none", "GPQA")
    gpqa_raw = Task("leaderboard_gpqa", "acc_norm,none", "GPQA Raw")

    musr = Task("leaderboard_musr", "acc_norm,none", "MUSR")
    musr_raw = Task("leaderboard_musr", "acc_norm,none", "MUSR Raw")

    mmlu_pro = Task("leaderboard_mmlu_pro", "acc,none", "MMLU-PRO")
    mmlu_pro_raw = Task("leaderboard_mmlu_pro", "acc,none", "MMLU-PRO Raw")


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass(frozen=True)
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    dummy: bool = False


auto_eval_column_dict = []
# Init
auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
# Scores
auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average ⬆️", "number", True)])
for task in Tasks:
    displayed_by_default = not task.name.endswith("_raw")
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", displayed_by_default=displayed_by_default)])
# Model information
auto_eval_column_dict.append(["model_type", ColumnContent, ColumnContent("Type", "str", False)])
auto_eval_column_dict.append(["architecture", ColumnContent, ColumnContent("Architecture", "str", False)])
auto_eval_column_dict.append(["weight_type", ColumnContent, ColumnContent("Weight type", "str", False, True)])
auto_eval_column_dict.append(["precision", ColumnContent, ColumnContent("Precision", "str", False)])
auto_eval_column_dict.append(["merged", ColumnContent, ColumnContent("Merged", "bool", False)])
auto_eval_column_dict.append(["license", ColumnContent, ColumnContent("Hub License", "str", False)])
auto_eval_column_dict.append(["params", ColumnContent, ColumnContent("#Params (B)", "number", False)])
auto_eval_column_dict.append(["likes", ColumnContent, ColumnContent("Hub ❀️", "number", False)])
auto_eval_column_dict.append(
    ["still_on_hub", ColumnContent, ColumnContent("Available on the hub", "bool", False, hidden=True)]
)
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
auto_eval_column_dict.append(["not_flagged", ColumnContent, ColumnContent("Flagged", "bool", False, hidden=True)])
auto_eval_column_dict.append(["moe", ColumnContent, ColumnContent("MoE", "bool", False, hidden=True)])
auto_eval_column_dict.append(["date", ColumnContent, ColumnContent("date", "bool", False, hidden=True)])
auto_eval_column_dict.append(["use_chat_template", ColumnContent, ColumnContent("Chat Template", "bool", False)])
auto_eval_column_dict.append(["maintainers_highlight", ColumnContent, ColumnContent("Maintainer's Highlight", "bool", False, hidden=True)])

# fullname structure: <user>/<model_name>
auto_eval_column_dict.append(["fullname", ColumnContent, ColumnContent("fullname", "str", False, dummy=True)])

# We use make dataclass to dynamically fill the scores from Tasks
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)


@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model_link = ColumnContent("model_link", "markdown", True)
    model_name = ColumnContent("model_name", "str", True)
    revision = ColumnContent("revision", "str", True)
    #private = ColumnContent("private", "bool", True)  # Should not be displayed
    precision = ColumnContent("precision", "str", True)
    #weight_type = ColumnContent("weight_type", "str", "Original") # Might be confusing, to think about
    status = ColumnContent("status", "str", True)


# baseline_row = {
#     AutoEvalColumn.model.name: "<p>Baseline</p>",
#     AutoEvalColumn.revision.name: "N/A",
#     AutoEvalColumn.precision.name: None,
#     AutoEvalColumn.merged.name: False,
#     AutoEvalColumn.average.name: 31.0,
#     AutoEvalColumn.arc.name: 25.0,
#     AutoEvalColumn.hellaswag.name: 25.0,
#     AutoEvalColumn.mmlu.name: 25.0,
#     AutoEvalColumn.truthfulqa.name: 25.0,
#     AutoEvalColumn.winogrande.name: 50.0,
#     AutoEvalColumn.gsm8k.name: 0.21,
#     AutoEvalColumn.fullname.name: "baseline",
#     AutoEvalColumn.model_type.name: "",
#     AutoEvalColumn.not_flagged.name: False,
# }

# Average ⬆️ human baseline is 0.897 (source: averaging human baselines below)
# ARC human baseline is 0.80 (source: https://lab42.global/arc/)
# HellaSwag human baseline is 0.95 (source: https://deepgram.com/learn/hellaswag-llm-benchmark-guide)
# MMLU human baseline is 0.898 (source: https://openreview.net/forum?id=d7KBjmI3GmQ)
# TruthfulQA human baseline is 0.94(source: https://arxiv.org/pdf/2109.07958.pdf)
# Winogrande: https://leaderboard.allenai.org/winogrande/submissions/public
# GSM8K: paper
# Define the human baselines
# human_baseline_row = {
#     AutoEvalColumn.model.name: "<p>Human performance</p>",
#     AutoEvalColumn.revision.name: "N/A",
#     AutoEvalColumn.precision.name: None,
#     AutoEvalColumn.average.name: 92.75,
#     AutoEvalColumn.merged.name: False,
#     AutoEvalColumn.arc.name: 80.0,
#     AutoEvalColumn.hellaswag.name: 95.0,
#     AutoEvalColumn.mmlu.name: 89.8,
#     AutoEvalColumn.truthfulqa.name: 94.0,
#     AutoEvalColumn.winogrande.name: 94.0,
#     AutoEvalColumn.gsm8k.name: 100,
#     AutoEvalColumn.fullname.name: "human_baseline",
#     AutoEvalColumn.model_type.name: "",
#     AutoEvalColumn.not_flagged.name: False,
# }


@dataclass
class ModelDetails:
    name: str
    symbol: str = ""  # emoji, only for the model type


class ModelType(Enum):
    PT = ModelDetails(name="🟒 pretrained", symbol="🟒")
    CPT = ModelDetails(name="🟩 continuously pretrained", symbol="🟩")
    FT = ModelDetails(name="πŸ”Ά fine-tuned on domain-specific datasets", symbol="πŸ”Ά")
    chat = ModelDetails(name="πŸ’¬ chat models (RLHF, DPO, IFT, ...)", symbol="πŸ’¬")
    merges = ModelDetails(name="🀝 base merges and moerges", symbol="🀝")
    Unknown = ModelDetails(name="❓ other", symbol="❓")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(m_type):
        if any([k for k in m_type if k in ["fine-tuned","πŸ”Ά", "finetuned"]]):
            return ModelType.FT
        if "continuously pretrained" in m_type or "🟩" in m_type:
            return ModelType.CPT
        if "pretrained" in m_type or "🟒" in m_type:
            return ModelType.PT
        if any([k in m_type for k in ["instruction-tuned", "RL-tuned", "chat", "🟦", "β­•", "πŸ’¬"]]):
            return ModelType.chat
        if "merge" in m_type or "🀝" in m_type:
            return ModelType.merges
        return ModelType.Unknown


class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")


class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    qt_8bit = ModelDetails("8bit")
    qt_4bit = ModelDetails("4bit")
    qt_GPTQ = ModelDetails("GPTQ")
    Unknown = ModelDetails("?")

    @staticmethod
    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["8bit"]:
            return Precision.qt_8bit
        if precision in ["4bit"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.Unknown


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn)]
TYPES = [c.type for c in fields(AutoEvalColumn)]

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

BENCHMARK_COLS = [t.value.col_name for t in Tasks]

NUMERIC_INTERVALS = {
    "?": pd.Interval(-1, 0, closed="right"),
    "~1.5": pd.Interval(0, 2, closed="right"),
    "~3": pd.Interval(2, 4, closed="right"),
    "~7": pd.Interval(4, 9, closed="right"),
    "~13": pd.Interval(9, 20, closed="right"),
    "~35": pd.Interval(20, 45, closed="right"),
    "~60": pd.Interval(45, 70, closed="right"),
    "70+": pd.Interval(70, 10000, closed="right"),
}