opencompass/examples/eval_subjective_bradleyterry.py at main · BudEcosystem/opencompass · GitHub

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
from mmengine.config import read_base

with read_base():
    from opencompass.configs.datasets.subjective.alpaca_eval.alpacav2_judgeby_gpt4_bradleyterry import (
        alpacav2_datasets, )

    from opencompass.configs.datasets.subjective.arena_hard.arena_hard_compare_bradleyterry import (
        arenahard_datasets, )

    from opencompass.configs.datasets.subjective.compassarena.compassarena_compare_bradleyterry import (
        compassarena_datasets, )

    from opencompass.configs.datasets.subjective.wildbench.wildbench_pair_judge_bradleyterry import (
        wildbench_datasets, )

    from opencompass.configs.models.hf_internlm.lmdeploy_internlm2_5_7b_chat import (
        models as lmdeploy_internlm2_5_7b_chat, )

    from opencompass.configs.models.hf_internlm.lmdeploy_internlm2_5_20b_chat import (
        models as lmdeploy_internlm2_5_20b_chat, )

    from opencompass.configs.models.qwen2_5.lmdeploy_qwen2_5_7b_instruct import (
        models as lmdeploy_qwen2_5_7b_instruct, )

    from opencompass.configs.models.qwen2_5.lmdeploy_qwen2_5_14b_instruct import (
        models as lmdeploy_qwen2_5_14b_instruct, )

    from opencompass.configs.models.qwen.lmdeploy_qwen2_7b_instruct import (
        models as lmdeploy_qwen2_7b_instruct, )

from opencompass.models import (HuggingFace, HuggingFaceCausalLM,
                                HuggingFaceChatGLM3, OpenAI,
                                TurboMindModelwithChatTemplate)
from opencompass.partitioners import NaivePartitioner, SizePartitioner
from opencompass.partitioners.sub_naive import SubjectiveNaivePartitioner
from opencompass.partitioners.sub_num_worker import \
    SubjectiveNumWorkerPartitioner
from opencompass.partitioners.sub_size import SubjectiveSizePartitioner
from opencompass.runners import LocalRunner, SlurmSequentialRunner
from opencompass.summarizers import (CompassArenaBradleyTerrySummarizer,
                                     SubjectiveSummarizer)
from opencompass.tasks import OpenICLInferTask
from opencompass.tasks.subjective_eval import SubjectiveEvalTask

api_meta_template = dict(round=[
    dict(role='HUMAN', api_role='HUMAN'),
    dict(role='BOT', api_role='BOT', generate=True),
])

# -------------Inference Stage ----------------------------------------
# For subjective evaluation, we often set do sample for models
models = [
    *lmdeploy_internlm2_5_7b_chat,
    *lmdeploy_internlm2_5_20b_chat,
    *lmdeploy_qwen2_5_14b_instruct,
    *lmdeploy_qwen2_5_7b_instruct,
    *lmdeploy_qwen2_7b_instruct,
]

datasets = [
    *alpacav2_datasets,
    *arenahard_datasets,
    *compassarena_datasets,
    *wildbench_datasets,
]

infer = dict(
    partitioner=dict(type=NaivePartitioner),
    runner=dict(type=LocalRunner,
                max_num_workers=16,
                task=dict(type=OpenICLInferTask)),
)
# -------------Evalation Stage ----------------------------------------

## ------------- JudgeLLM Configuration
judge_models = [
    dict(
        type=TurboMindModelwithChatTemplate,
        abbr='CompassJudger-1-32B-Instruct',
        path='opencompass/CompassJudger-1-32B-Instruct',
        engine_config=dict(session_len=16384, max_batch_size=16, tp=4),
        gen_config=dict(top_k=1,
                        temperature=1e-6,
                        top_p=0.9,
                        max_new_tokens=2048),
        max_seq_len=16384,
        max_out_len=2048,
        batch_size=16,
        run_cfg=dict(num_gpus=4),
    )
]

## ------------- Evaluation Configuration
eval = dict(
    partitioner=dict(
        type=SubjectiveNaivePartitioner,
        models=models,
        judge_models=judge_models,
    ),
    runner=dict(type=LocalRunner,
                max_num_workers=16,
                task=dict(type=SubjectiveEvalTask)),
)

## ------------- Summary Configuration
# This step fits a Bradley-Terry model (statistical model) with an option
# to include style features and control variables based on groups
# (group variables must be available in the input dataset for each observation).
summarizer = dict(
    type=CompassArenaBradleyTerrySummarizer,
    rating_system='bradleyterry',
    report_pred_win_rates=True,
    num_bootstrap=100,
    num_cpu=None,
    with_control_vars=True,
    normalize_style_features=False,
    odds_ratio=True,
)

work_dir = 'outputs/subjective/bradleyterry'