File size: 7,558 Bytes
018441b
a8ede2f
 
 
 
dc1ba50
018441b
 
 
 
 
 
 
 
 
 
43578e7
018441b
0b743c4
ba4f18b
 
a774132
7de3b23
 
 
a774132
7de3b23
 
 
a774132
7de3b23
 
 
a774132
7de3b23
bfbf195
a774132
7de3b23
 
a774132
7de3b23
 
 
 
 
b3ee622
e5e2b84
a3e58c8
 
1ff7ef9
a3e58c8
 
2e5fa57
1ff7ef9
e5e2b84
a8ede2f
 
 
 
 
 
 
 
 
 
 
 
 
018441b
 
 
 
7d1a89f
018441b
7d1a89f
 
5e86dd4
 
 
018441b
 
7d1a89f
018441b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8ede2f
61c2746
a8ede2f
 
 
 
 
 
 
 
 
 
 
018441b
a8ede2f
dc1ba50
a8ede2f
 
 
018441b
fdb7c69
 
 
018441b
a8ede2f
 
 
 
 
 
 
 
 
 
fdb7c69
 
 
 
a8ede2f
 
dc1ba50
018441b
 
 
 
 
dc1ba50
018441b
dc1ba50
018441b
 
 
 
 
 
 
dc1ba50
 
 
 
018441b
 
 
 
 
 
 
 
 
 
 
69021cc
a8ede2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from dataclasses import dataclass, make_dataclass
from enum import Enum

import pandas as pd


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):
    # XXX include me back at some point
    nqopen = Task("nq8", "em", "NQ Open/EM")
    triviaqa = Task("tqa8", "em", "TriviaQA/EM")

    truthfulqa_mc1 = Task("truthfulqa_mc1", "acc", "TruthQA MC1/Acc")
    truthfulqa_mc2 = Task("truthfulqa_mc2", "acc", "TruthQA MC2/Acc")
    truthfulqa_gen = Task("truthfulqa_gen", "rougeL_acc", "TruthQA Gen/ROUGE")

    xsum_r = Task("xsum_v2", "rougeL", "XSum/ROUGE")
    xsum_f = Task("xsum_v2", "factKB", "XSum/factKB")
    xsum_b = Task("xsum_v2", "bertscore_precision", "XSum/BERT-P")

    cnndm_r = Task("cnndm_v2", "rougeL", "CNN-DM/ROUGE")
    cnndm_f = Task("cnndm_v2", "factKB", "CNN-DM/factKB")
    cnndm_b = Task("cnndm_v2", "bertscore_precision", "CNN-DM/BERT-P")

    race = Task("race", "acc", "RACE/Acc")
    squadv2 = Task("squadv2", "exact", "SQuADv2/EM")

    memotrap = Task("memo-trap_v2", "acc", "MemoTrap/Acc")
    ifeval = Task("ifeval", "prompt_level_strict_acc", "IFEval/Acc")

    faithdial = Task("faithdial_hallu_v2", "acc", "FaithDial/Acc")

    halueval_qa = Task("halueval_qa", "acc", "HaluQA/Acc")
    halueval_summ = Task("halueval_summarization", "acc", "HaluSumm/Acc")
    halueval_dial = Task("halueval_dialogue", "acc", "HaluDial/Acc")

    # XXX include me back at some point
    #selfcheck = Task("selfcheckgpt", "max-selfcheckgpt", "SelfCheckGPT")

    # add popqa, fever, nq-swap and truefalse
    fever11 = Task("fever11", "acc", "FEVER/Acc")
    truefalse_cieacf = Task("truefalse_cieacf", "acc", "TrueFalse/Acc")
    popqa = Task("popqa", "em", "PopQA/EM")
    nq_swap = Task("nq_swap", "em", "NQ-Swap/EM")


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
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("Avg", "number", True)])

auto_eval_column_dict.append(["Faithfulness", ColumnContent, ColumnContent("Faithfulness", "number", True)])
auto_eval_column_dict.append(["Factuality", ColumnContent, ColumnContent("Factuality", "number", True)])

for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True)])

# 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(["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)])
auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, False)])
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent, ColumnContent("model_name_for_query", "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 = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    private = ColumnContent("private", "bool", True)
    precision = ColumnContent("precision", "str", True)
    weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


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


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

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

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


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


class Precision(Enum):
    float32 = ModelDetails("float32")
    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: str):
        if precision in ["torch.float32", "float32"]:
            return Precision.float32
        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) 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]

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"),
}