style: pass all formatting hooks (yapf & quote fixer)

This commit is contained in:
root 2025-04-28 08:03:47 +00:00
parent 8c74e6a39e
commit e2f80574ec
11 changed files with 967 additions and 5 deletions

View File

@ -20,7 +20,7 @@ exclude: |
)
repos:
- repo: https://github.com/PyCQA/flake8
rev: 5.0.4
rev: 7.2.0
hooks:
- id: flake8
exclude: |
@ -29,7 +29,7 @@ repos:
examples/
)
- repo: https://github.com/PyCQA/isort
rev: 5.11.5
rev: 6.0.1
hooks:
- id: isort
exclude: |
@ -47,7 +47,7 @@ repos:
examples/
)
- repo: https://github.com/codespell-project/codespell
rev: v2.2.1
rev: v2.4.1
hooks:
- id: codespell
exclude: |
@ -58,7 +58,7 @@ repos:
examples/
)
- repo: https://github.com/pre-commit/pre-commit-hooks
rev: v4.3.0
rev: v5.0.0
hooks:
- id: trailing-whitespace
exclude: |
@ -81,7 +81,7 @@ repos:
- id: mixed-line-ending
args: ["--fix=lf"]
- repo: https://github.com/executablebooks/mdformat
rev: 0.7.9
rev: 0.7.22
hooks:
- id: mdformat
args: ["--number", "--table-width", "200"]

View File

@ -0,0 +1,61 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import FixKRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.openicl.icl_evaluator import AccwithDetailsEvaluator
from opencompass.utils.text_postprocessors import first_option_postprocess
from opencompass.datasets import SciEvalDataset # 你自己实现的类
# 只评测 biology + multiple-choice 的 test split
_hint = ('Given a question and four options, please select the right answer. '
"Your answer should be 'A', 'B', 'C' or 'D'.")
scieval_reader_cfg = dict(
input_columns=['input', 'A', 'B', 'C', 'D'],
output_column='target',
train_split='test',
)
scieval_infer_cfg = dict(
ice_template=dict(
type=PromptTemplate,
template=dict(round=[
dict(
role='HUMAN',
prompt=f'{_hint}\nQuestion: {{input}}\nA. {{A}}\nB. {{B}}\nC. {{C}}\nD. {{D}}\nAnswer: '
),
dict(role='BOT', prompt='{target}\n')
]),
),
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin='</E>',
round=[
dict(
role='HUMAN',
prompt=f'{_hint}\nQuestion: {{input}}\nA. {{A}}\nB. {{B}}\nC. {{C}}\nD. {{D}}\nAnswer: '
),
],
),
ice_token='</E>',
),
retriever=dict(type=FixKRetriever, fix_id_list=[0, 1, 2, 3, 4]),
inferencer=dict(type=GenInferencer),
)
scieval_eval_cfg = dict(
evaluator=dict(type=AccwithDetailsEvaluator),
pred_postprocessor=dict(type=first_option_postprocess, options='ABCD'),
)
scieval_datasets = [
dict(
abbr='scieval_biology',
type=SciEvalDataset,
path='OpenDFM/SciEval',
name='default',
reader_cfg=scieval_reader_cfg,
infer_cfg=scieval_infer_cfg,
eval_cfg=scieval_eval_cfg,
)
]

View File

@ -0,0 +1,125 @@
# SciEval_lifescience_llmjudge_gen.py
from mmengine.config import read_base
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.utils.text_postprocessors import match_answer_pattern
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
from opencompass.datasets import SciEvalDataset
with read_base():
from .SciEval_lifescience_sets import SciEval_lifescience_subsets
QUERY_TEMPLATE = """
Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: $LETTER' (without quotes) where LETTER is one of ABCD.
{input}
A) {A}
B) {B}
C) {C}
D) {D}
""".strip()
GRADER_TEMPLATE = """
Please as a grading expert, judge whether the final answers given by the candidates below are consistent with the standard answers, that is, whether the candidates answered correctly.
Here are some evaluation criteria:
1. Please refer to the given standard answer. You don't need to re-generate the answer to the question because the standard answer has been given. You only need to judge whether the candidate's answer is consistent with the standard answer according to the form of the question. Don't try to answer the original question. You can assume that the standard answer is definitely correct.
2. Because the candidate's answer may be different from the standard answer in the form of expression, before making a judgment, please understand the question and the standard answer first, and then judge whether the candidate's answer is correct, but be careful not to try to answer the original question.
3. Some answers may contain multiple items, such as multiple-choice questions, multiple-select questions, fill-in-the-blank questions, etc. As long as the answer is the same as the standard answer, it is enough. For multiple-select questions and multiple-blank fill-in-the-blank questions, the candidate needs to answer all the corresponding options or blanks correctly to be considered correct.
4. Some answers may be expressed in different ways, such as some answers may be a mathematical expression, some answers may be a textual description, as long as the meaning expressed is the same. And some formulas are expressed in different ways, but they are equivalent and correct.
Please judge whether the following answers are consistent with the standard answer based on the above criteria. Grade the predicted answer of this new question as one of:
A: CORRECT
B: INCORRECT
Just return the letters "A" or "B", with no text around it.
Here is your task. Simply reply with either CORRECT, INCORRECT. Don't apologize or correct yourself if there was a mistake; we are just trying to grade the answer.
<Original Question Begin>: {input}
A) {A}
B) {B}
C) {C}
D) {D}
<Original Question End>
<Gold Target Begin>:
{target}
<Gold Target End>
<Predicted Answer Begin>:
{prediction}
<Predicted End>
Judging the correctness of candidates' answers:
""".strip()
scieval_reader_cfg = dict(
input_columns=['input', 'A', 'B', 'C', 'D'],
output_column='target',
train_split='test',
)
scieval_datasets = []
for name in SciEval_lifescience_subsets:
scieval_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt=QUERY_TEMPLATE),
]
)
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
scieval_eval_cfg = dict(
evaluator=dict(
type=GenericLLMEvaluator,
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin=[
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt=(
'You are a helpful assistant who evaluates the correctness '
"and quality of models' outputs."
),
)
],
round=[
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=SciEvalDataset,
path='OpenDFM/SciEval',
name='default',
reader_cfg=scieval_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
scieval_datasets.append(
dict(
abbr=f'scieval_lifescience_{name}_llmjudge',
type=SciEvalDataset,
path='OpenDFM/SciEval',
name='default',
reader_cfg=scieval_reader_cfg,
infer_cfg=scieval_infer_cfg,
eval_cfg=scieval_eval_cfg,
mode='singlescore',
)
)

View File

@ -0,0 +1,3 @@
SciEval_lifescience_subsets = [
'biology', # 大学生物学
]

View File

@ -0,0 +1,368 @@
# MMLU
```bash
python3 run.py --models hf_internlm2_7b --datasets mmlu_ppl_ac766d --debug
python3 run.py --models hf_internlm2_chat_7b --datasets mmlu_gen_4d595a --debug
```
## Base Models
| model | mmlu | mmlu-stem | mmlu-social-science | mmlu-humanities | mmlu-other |
|:------------------------:|-------:|------------:|----------------------:|------------------:|-------------:|
| llama-7b-turbomind | 35.66 | 31.22 | 37.70 | 38.90 | 37.01 |
| llama-13b-turbomind | 47.76 | 37.68 | 55.36 | 52.43 | 50.83 |
| llama-30b-turbomind | 58.55 | 46.95 | 67.35 | 65.13 | 60.78 |
| llama-65b-turbomind | 63.78 | 52.35 | 73.68 | 70.84 | 64.29 |
| llama-2-7b-turbomind | 46.78 | 37.81 | 52.11 | 51.69 | 50.04 |
| llama-2-13b-turbomind | 55.76 | 44.61 | 63.86 | 62.97 | 57.35 |
| llama-2-70b-turbomind | 69.87 | 58.30 | 79.86 | 75.84 | 71.58 |
| llama-3-8b-turbomind | 66.43 | 55.95 | 76.11 | 70.29 | 68.96 |
| llama-3-70b-turbomind | 79.35 | 70.66 | 87.54 | 83.43 | 80.42 |
| internlm2-1.8b-turbomind | 45.99 | 39.63 | 51.02 | 48.65 | 47.96 |
| internlm2-7b-turbomind | 65.84 | 56.48 | 74.43 | 69.68 | 67.75 |
| internlm2-20b-turbomind | 67.58 | 59.01 | 76.04 | 71.20 | 68.69 |
| qwen-1.8b-turbomind | 46.61 | 38.91 | 51.35 | 49.57 | 50.51 |
| qwen-7b-turbomind | 59.75 | 50.16 | 67.98 | 63.48 | 62.44 |
| qwen-14b-turbomind | 67.85 | 59.13 | 76.18 | 71.62 | 69.12 |
| qwen-72b-turbomind | 77.36 | 68.70 | 85.28 | 80.60 | 79.45 |
| qwen1.5-0.5b-hf | 39.98 | 33.96 | 45.08 | 41.59 | 42.48 |
| qwen1.5-1.8b-hf | 47.14 | 39.47 | 52.70 | 49.01 | 51.33 |
| qwen1.5-4b-hf | 57.03 | 47.80 | 64.86 | 60.10 | 60.20 |
| qwen1.5-7b-hf | 62.15 | 53.22 | 70.25 | 65.62 | 64.26 |
| qwen1.5-14b-hf | 69.10 | 61.46 | 77.57 | 71.25 | 70.29 |
| qwen1.5-32b-hf | 73.88 | 65.60 | 81.41 | 77.10 | 75.79 |
| qwen1.5-72b-hf | 77.02 | 69.00 | 84.55 | 80.60 | 78.21 |
| qwen1.5-moe-a2-7b-hf | 62.09 | 53.27 | 70.74 | 63.80 | 65.28 |
| mistral-7b-v0.1-hf | 64.04 | 53.21 | 73.65 | 68.04 | 67.00 |
| mistral-7b-v0.2-hf | 63.85 | 53.21 | 72.17 | 68.40 | 67.15 |
| mixtral-8x7b-v0.1-hf | 71.80 | 61.70 | 81.03 | 75.51 | 74.35 |
| mixtral-8x22b-v0.1-hf | 77.67 | 68.94 | 86.81 | 81.23 | 78.43 |
| yi-6b-hf | 64.08 | 52.61 | 74.10 | 68.58 | 67.11 |
| yi-34b-hf | 76.26 | 66.73 | 83.74 | 81.78 | 77.77 |
| deepseek-7b-base-hf | 49.22 | 40.17 | 56.73 | 53.46 | 51.26 |
| deepseek-67b-base-hf | 71.95 | 60.57 | 81.69 | 77.11 | 74.42 |
### Details
| model | college_biology | college_chemistry | college_computer_science | college_mathematics | college_physics | electrical_engineering | astronomy | anatomy | abstract_algebra | machine_learning | clinical_knowledge | global_facts |
|:------------------------:|------------------:|--------------------:|---------------------------:|----------------------:|------------------:|-------------------------:|------------:|----------:|-------------------:|-------------------:|---------------------:|---------------:|
| llama-7b-turbomind | 37.50 | 30.00 | 30.00 | 33.00 | 23.53 | 23.45 | 34.87 | 37.78 | 25.00 | 27.68 | 34.34 | 31.00 |
| llama-13b-turbomind | 46.53 | 30.00 | 42.00 | 36.00 | 18.63 | 42.76 | 46.71 | 46.67 | 30.00 | 32.14 | 45.66 | 37.00 |
| llama-30b-turbomind | 59.03 | 45.00 | 47.00 | 35.00 | 26.47 | 53.10 | 61.18 | 51.85 | 37.00 | 41.07 | 57.36 | 38.00 |
| llama-65b-turbomind | 68.75 | 49.00 | 47.00 | 37.00 | 35.29 | 55.17 | 73.03 | 57.78 | 30.00 | 48.21 | 66.04 | 38.00 |
| llama-2-7b-turbomind | 46.53 | 34.00 | 33.00 | 34.00 | 22.55 | 47.59 | 40.13 | 47.41 | 29.00 | 38.39 | 46.42 | 32.00 |
| llama-2-13b-turbomind | 59.03 | 44.00 | 48.00 | 29.00 | 26.47 | 50.34 | 53.29 | 49.63 | 35.00 | 28.57 | 60.00 | 32.00 |
| llama-2-70b-turbomind | 84.72 | 51.00 | 60.00 | 39.00 | 37.25 | 65.52 | 81.58 | 63.70 | 32.00 | 52.68 | 72.08 | 46.00 |
| llama-3-8b-turbomind | 77.08 | 46.00 | 51.00 | 31.00 | 51.96 | 62.76 | 67.11 | 68.15 | 34.00 | 52.68 | 74.72 | 35.00 |
| llama-3-70b-turbomind | 93.75 | 62.00 | 72.00 | 52.00 | 50.98 | 74.48 | 92.11 | 79.26 | 48.00 | 63.39 | 86.42 | 49.00 |
| internlm2-1.8b-turbomind | 38.89 | 37.00 | 44.00 | 35.00 | 30.39 | 49.66 | 50.66 | 44.44 | 25.00 | 35.71 | 51.32 | 32.00 |
| internlm2-7b-turbomind | 77.08 | 48.00 | 64.00 | 33.00 | 47.06 | 63.45 | 73.68 | 57.78 | 37.00 | 45.54 | 69.81 | 35.00 |
| internlm2-20b-turbomind | 83.33 | 51.00 | 61.00 | 36.00 | 45.10 | 64.83 | 75.00 | 59.26 | 39.00 | 53.57 | 73.58 | 32.00 |
| qwen-1.8b-turbomind | 42.36 | 36.00 | 39.00 | 34.00 | 27.45 | 51.03 | 50.66 | 42.96 | 31.00 | 31.25 | 53.21 | 28.00 |
| qwen-7b-turbomind | 67.36 | 48.00 | 53.00 | 28.00 | 39.22 | 59.31 | 63.82 | 49.63 | 34.00 | 38.39 | 63.02 | 37.00 |
| qwen-14b-turbomind | 78.47 | 51.00 | 62.00 | 42.00 | 49.02 | 65.52 | 71.05 | 60.00 | 37.00 | 58.93 | 71.32 | 40.00 |
| qwen-72b-turbomind | 93.75 | 56.00 | 66.00 | 56.00 | 50.98 | 80.69 | 85.53 | 73.33 | 41.00 | 62.50 | 83.77 | 54.00 |
| qwen1.5-0.5b-hf | 38.89 | 25.00 | 38.00 | 32.00 | 25.49 | 45.52 | 44.74 | 33.33 | 30.00 | 39.29 | 38.11 | 39.00 |
| qwen1.5-1.8b-hf | 43.75 | 34.00 | 45.00 | 38.00 | 28.43 | 47.59 | 47.37 | 40.74 | 32.00 | 31.25 | 53.96 | 37.00 |
| qwen1.5-4b-hf | 50.00 | 46.00 | 41.00 | 45.00 | 31.37 | 53.10 | 61.18 | 51.85 | 35.00 | 44.64 | 60.38 | 37.00 |
| qwen1.5-7b-hf | 66.67 | 48.00 | 55.00 | 37.00 | 41.18 | 60.69 | 65.79 | 52.59 | 39.00 | 41.07 | 68.68 | 43.00 |
| qwen1.5-14b-hf | 75.69 | 49.00 | 58.00 | 49.00 | 49.02 | 71.72 | 73.03 | 65.93 | 39.00 | 52.68 | 73.96 | 49.00 |
| qwen1.5-32b-hf | 85.42 | 53.00 | 59.00 | 51.00 | 53.92 | 72.41 | 82.24 | 63.70 | 43.00 | 58.04 | 78.11 | 50.00 |
| qwen1.5-72b-hf | 90.97 | 54.00 | 65.00 | 57.00 | 52.94 | 80.00 | 87.50 | 73.33 | 43.00 | 64.29 | 81.89 | 50.00 |
| qwen1.5-moe-a2-7b-hf | 62.50 | 44.00 | 54.00 | 41.00 | 49.02 | 58.62 | 69.74 | 57.78 | 37.00 | 38.39 | 66.79 | 38.00 |
| mistral-7b-v0.1-hf | 72.92 | 50.00 | 51.00 | 40.00 | 39.22 | 57.93 | 65.79 | 62.96 | 29.00 | 49.11 | 69.43 | 36.00 |
| mistral-7b-v0.2-hf | 71.53 | 49.00 | 53.00 | 40.00 | 36.27 | 57.24 | 64.47 | 60.00 | 29.00 | 53.57 | 67.92 | 39.00 |
| mixtral-8x7b-v0.1-hf | 85.42 | 54.00 | 62.00 | 43.00 | 46.08 | 68.97 | 82.89 | 70.37 | 37.00 | 56.25 | 79.25 | 51.00 |
| mixtral-8x22b-v0.1-hf | 89.58 | 56.00 | 69.00 | 48.00 | 52.94 | 76.55 | 86.18 | 77.04 | 53.00 | 62.50 | 82.26 | 56.00 |
| yi-6b-hf | 66.67 | 43.00 | 51.00 | 39.00 | 35.29 | 64.83 | 65.79 | 60.00 | 29.00 | 41.96 | 66.79 | 46.00 |
| yi-34b-hf | 88.89 | 52.00 | 66.00 | 44.00 | 48.04 | 80.00 | 89.47 | 74.81 | 44.00 | 58.04 | 78.87 | 52.00 |
| deepseek-7b-base-hf | 52.08 | 29.00 | 44.00 | 40.00 | 31.37 | 44.83 | 51.97 | 40.74 | 27.00 | 32.14 | 53.58 | 31.00 |
| deepseek-67b-base-hf | 84.72 | 52.00 | 62.00 | 42.00 | 42.16 | 70.34 | 80.92 | 65.19 | 39.00 | 50.00 | 78.11 | 42.00 |
| model | management | nutrition | marketing | professional_accounting | high_school_geography | international_law | moral_scenarios | computer_security | high_school_microeconomics | professional_law | medical_genetics | professional_psychology |
|:------------------------:|-------------:|------------:|------------:|--------------------------:|------------------------:|--------------------:|------------------:|--------------------:|-----------------------------:|-------------------:|-------------------:|--------------------------:|
| llama-7b-turbomind | 33.01 | 39.22 | 45.73 | 26.24 | 33.33 | 51.24 | 24.25 | 45.00 | 31.09 | 30.05 | 37.00 | 35.13 |
| llama-13b-turbomind | 66.02 | 51.63 | 71.79 | 34.75 | 55.05 | 64.46 | 30.06 | 63.00 | 47.48 | 37.22 | 53.00 | 48.53 |
| llama-30b-turbomind | 76.70 | 62.42 | 84.19 | 44.68 | 71.72 | 75.21 | 40.56 | 66.00 | 57.98 | 46.48 | 66.00 | 63.73 |
| llama-65b-turbomind | 82.52 | 68.95 | 87.18 | 48.94 | 79.29 | 81.82 | 47.82 | 79.00 | 68.49 | 50.07 | 68.00 | 66.67 |
| llama-2-7b-turbomind | 53.40 | 48.69 | 68.38 | 36.52 | 49.49 | 65.29 | 24.02 | 60.00 | 44.12 | 36.31 | 55.00 | 43.79 |
| llama-2-13b-turbomind | 72.82 | 61.76 | 79.49 | 39.72 | 69.19 | 74.38 | 43.80 | 70.00 | 58.40 | 42.50 | 54.00 | 54.90 |
| llama-2-70b-turbomind | 83.50 | 77.12 | 91.03 | 56.03 | 86.87 | 87.60 | 44.69 | 77.00 | 77.31 | 52.93 | 74.00 | 75.65 |
| llama-3-8b-turbomind | 87.38 | 75.82 | 89.74 | 48.94 | 80.81 | 84.30 | 40.89 | 81.00 | 73.95 | 46.22 | 77.00 | 71.90 |
| llama-3-70b-turbomind | 91.26 | 87.25 | 94.87 | 64.18 | 93.94 | 89.26 | 62.91 | 83.00 | 87.82 | 61.80 | 90.00 | 85.78 |
| internlm2-1.8b-turbomind | 60.19 | 58.17 | 63.25 | 31.21 | 56.57 | 56.20 | 24.47 | 52.00 | 50.42 | 36.11 | 53.00 | 41.83 |
| internlm2-7b-turbomind | 79.61 | 75.49 | 87.61 | 48.23 | 82.83 | 77.69 | 49.39 | 74.00 | 72.27 | 47.65 | 73.00 | 65.03 |
| internlm2-20b-turbomind | 79.61 | 75.49 | 91.88 | 50.00 | 87.88 | 85.95 | 35.08 | 81.00 | 70.59 | 49.48 | 78.00 | 70.10 |
| qwen-1.8b-turbomind | 66.02 | 60.46 | 73.50 | 38.30 | 56.57 | 66.94 | 23.91 | 56.00 | 42.02 | 33.96 | 51.00 | 39.54 |
| qwen-7b-turbomind | 78.64 | 67.32 | 83.33 | 41.49 | 76.77 | 76.03 | 29.72 | 73.00 | 58.40 | 41.72 | 69.00 | 59.64 |
| qwen-14b-turbomind | 78.64 | 73.86 | 88.89 | 48.58 | 83.84 | 84.30 | 45.47 | 77.00 | 73.95 | 50.85 | 74.00 | 69.61 |
| qwen-72b-turbomind | 90.29 | 84.97 | 94.87 | 65.96 | 92.93 | 88.43 | 65.70 | 79.00 | 84.87 | 61.21 | 86.00 | 82.19 |
| qwen1.5-0.5b-hf | 52.43 | 46.41 | 60.68 | 31.21 | 46.46 | 56.20 | 25.70 | 46.00 | 37.39 | 32.79 | 46.00 | 37.75 |
| qwen1.5-1.8b-hf | 66.02 | 58.50 | 75.64 | 33.69 | 56.06 | 72.73 | 24.69 | 57.00 | 39.50 | 36.11 | 53.00 | 42.81 |
| qwen1.5-4b-hf | 74.76 | 62.75 | 84.19 | 46.81 | 76.77 | 71.07 | 25.03 | 67.00 | 55.04 | 41.33 | 64.00 | 56.05 |
| qwen1.5-7b-hf | 78.64 | 70.92 | 86.32 | 44.68 | 81.82 | 77.69 | 32.74 | 76.00 | 64.29 | 45.37 | 68.00 | 61.27 |
| qwen1.5-14b-hf | 80.58 | 75.49 | 85.90 | 51.06 | 86.36 | 80.99 | 45.03 | 80.00 | 76.47 | 48.57 | 78.00 | 69.61 |
| qwen1.5-32b-hf | 86.41 | 81.37 | 95.30 | 56.38 | 91.41 | 88.43 | 44.02 | 76.00 | 82.77 | 57.89 | 83.00 | 75.33 |
| qwen1.5-72b-hf | 87.38 | 85.29 | 94.87 | 64.89 | 92.42 | 90.08 | 62.12 | 83.00 | 84.03 | 60.76 | 86.00 | 81.05 |
| qwen1.5-moe-a2-7b-hf | 78.64 | 70.92 | 86.32 | 46.81 | 81.82 | 77.69 | 25.59 | 71.00 | 65.97 | 45.37 | 65.00 | 61.44 |
| mistral-7b-v0.1-hf | 82.52 | 75.49 | 87.61 | 48.94 | 76.77 | 77.69 | 32.51 | 77.00 | 66.39 | 44.98 | 74.00 | 67.97 |
| mistral-7b-v0.2-hf | 81.55 | 74.18 | 88.46 | 51.06 | 76.77 | 80.99 | 38.77 | 75.00 | 64.71 | 45.37 | 72.00 | 66.34 |
| mixtral-8x7b-v0.1-hf | 87.38 | 81.70 | 91.88 | 51.77 | 85.86 | 85.95 | 40.11 | 80.00 | 79.41 | 53.32 | 77.00 | 77.94 |
| mixtral-8x22b-v0.1-hf | 89.32 | 85.95 | 91.88 | 62.06 | 91.41 | 90.08 | 64.58 | 83.00 | 87.82 | 60.82 | 84.00 | 83.17 |
| yi-6b-hf | 80.58 | 71.57 | 91.03 | 48.23 | 83.33 | 76.86 | 41.34 | 75.00 | 74.79 | 49.35 | 80.00 | 65.69 |
| yi-34b-hf | 91.26 | 85.62 | 92.31 | 65.25 | 89.39 | 91.74 | 64.69 | 82.00 | 85.29 | 59.97 | 87.00 | 82.19 |
| deepseek-7b-base-hf | 61.17 | 53.59 | 72.22 | 34.04 | 59.09 | 65.29 | 26.37 | 61.00 | 44.96 | 35.53 | 56.00 | 49.18 |
| deepseek-67b-base-hf | 88.35 | 79.74 | 91.88 | 57.09 | 89.39 | 85.12 | 46.15 | 76.00 | 82.35 | 55.93 | 72.00 | 79.58 |
| model | jurisprudence | world_religions | philosophy | virology | high_school_chemistry | public_relations | high_school_macroeconomics | human_sexuality | elementary_mathematics | high_school_physics | high_school_computer_science | high_school_european_history |
|:------------------------:|----------------:|------------------:|-------------:|-----------:|------------------------:|-------------------:|-----------------------------:|------------------:|-------------------------:|----------------------:|-------------------------------:|-------------------------------:|
| llama-7b-turbomind | 41.67 | 49.12 | 40.84 | 34.94 | 29.56 | 40.00 | 34.10 | 35.11 | 26.46 | 27.81 | 34.00 | 41.82 |
| llama-13b-turbomind | 51.85 | 67.84 | 55.31 | 43.37 | 28.57 | 60.91 | 46.15 | 57.25 | 26.98 | 29.80 | 49.00 | 61.21 |
| llama-30b-turbomind | 71.30 | 79.53 | 66.24 | 49.40 | 40.39 | 70.00 | 56.67 | 64.89 | 37.30 | 35.10 | 60.00 | 70.91 |
| llama-65b-turbomind | 75.00 | 81.29 | 73.63 | 53.01 | 41.38 | 74.55 | 65.90 | 77.86 | 40.21 | 35.76 | 69.00 | 76.36 |
| llama-2-7b-turbomind | 53.70 | 69.01 | 60.13 | 41.57 | 36.95 | 54.55 | 45.90 | 55.73 | 27.25 | 31.13 | 40.00 | 59.39 |
| llama-2-13b-turbomind | 74.07 | 76.61 | 63.99 | 45.78 | 44.83 | 62.73 | 50.77 | 62.60 | 34.13 | 36.42 | 57.00 | 63.03 |
| llama-2-70b-turbomind | 83.33 | 85.96 | 78.46 | 53.61 | 52.22 | 69.09 | 74.87 | 87.02 | 43.39 | 43.71 | 78.00 | 84.24 |
| llama-3-8b-turbomind | 75.00 | 83.04 | 74.28 | 56.02 | 54.68 | 71.82 | 64.87 | 79.39 | 42.06 | 45.03 | 68.00 | 76.36 |
| llama-3-70b-turbomind | 86.11 | 91.23 | 86.50 | 57.83 | 71.92 | 74.55 | 82.56 | 88.55 | 62.70 | 56.95 | 86.00 | 86.67 |
| internlm2-1.8b-turbomind | 55.56 | 59.65 | 51.13 | 40.96 | 43.35 | 52.73 | 43.33 | 47.33 | 30.42 | 33.11 | 47.00 | 56.36 |
| internlm2-7b-turbomind | 79.63 | 82.46 | 73.63 | 51.20 | 55.17 | 70.00 | 66.92 | 70.99 | 46.03 | 42.38 | 70.00 | 78.79 |
| internlm2-20b-turbomind | 75.93 | 82.46 | 73.95 | 56.02 | 57.64 | 68.18 | 70.51 | 68.70 | 49.21 | 38.41 | 75.00 | 82.42 |
| qwen-1.8b-turbomind | 59.26 | 56.14 | 50.80 | 40.96 | 37.93 | 60.00 | 41.03 | 51.15 | 33.33 | 34.44 | 39.00 | 64.24 |
| qwen-7b-turbomind | 73.15 | 76.61 | 67.20 | 47.59 | 51.23 | 65.45 | 60.00 | 69.47 | 43.12 | 38.41 | 67.00 | 66.67 |
| qwen-14b-turbomind | 76.85 | 84.21 | 72.03 | 53.01 | 65.52 | 66.36 | 66.92 | 78.63 | 51.32 | 41.72 | 72.00 | 82.42 |
| qwen-72b-turbomind | 83.33 | 88.30 | 83.28 | 58.43 | 65.52 | 74.55 | 81.54 | 89.31 | 68.52 | 58.28 | 81.00 | 84.24 |
| qwen1.5-0.5b-hf | 40.74 | 40.94 | 41.48 | 40.96 | 28.57 | 50.91 | 36.92 | 41.98 | 28.84 | 22.52 | 37.00 | 52.73 |
| qwen1.5-1.8b-hf | 55.56 | 57.31 | 49.84 | 40.96 | 36.45 | 56.36 | 43.59 | 56.49 | 35.19 | 27.81 | 45.00 | 61.21 |
| qwen1.5-4b-hf | 70.37 | 70.76 | 61.74 | 44.58 | 45.32 | 65.45 | 54.62 | 64.89 | 47.88 | 32.45 | 62.00 | 70.30 |
| qwen1.5-7b-hf | 75.93 | 77.19 | 66.24 | 50.60 | 53.20 | 62.73 | 60.00 | 71.76 | 50.26 | 38.41 | 71.00 | 74.55 |
| qwen1.5-14b-hf | 74.07 | 83.63 | 70.74 | 46.39 | 58.62 | 64.55 | 73.59 | 76.34 | 59.26 | 49.01 | 75.00 | 83.64 |
| qwen1.5-32b-hf | 83.33 | 85.96 | 82.96 | 56.63 | 61.58 | 63.64 | 77.95 | 83.97 | 69.31 | 50.99 | 85.00 | 86.06 |
| qwen1.5-72b-hf | 84.26 | 88.89 | 82.32 | 57.23 | 66.01 | 72.73 | 82.05 | 87.02 | 69.31 | 56.95 | 84.00 | 84.24 |
| qwen1.5-moe-a2-7b-hf | 70.37 | 80.12 | 66.56 | 51.20 | 47.78 | 64.55 | 62.31 | 70.99 | 46.30 | 45.03 | 59.00 | 69.70 |
| mistral-7b-v0.1-hf | 77.78 | 83.04 | 69.45 | 54.82 | 53.20 | 67.27 | 66.15 | 78.63 | 38.10 | 31.79 | 68.00 | 78.79 |
| mistral-7b-v0.2-hf | 73.15 | 82.46 | 72.99 | 53.01 | 55.67 | 66.36 | 62.31 | 77.10 | 40.48 | 34.44 | 66.00 | 76.36 |
| mixtral-8x7b-v0.1-hf | 82.41 | 88.30 | 78.14 | 51.20 | 62.56 | 70.00 | 70.77 | 80.92 | 48.68 | 48.34 | 71.00 | 80.61 |
| mixtral-8x22b-v0.1-hf | 84.26 | 89.47 | 84.57 | 59.04 | 67.49 | 78.18 | 79.23 | 88.55 | 61.64 | 52.98 | 87.00 | 86.06 |
| yi-6b-hf | 78.70 | 81.87 | 69.77 | 46.39 | 52.71 | 73.64 | 65.13 | 74.81 | 46.30 | 38.41 | 66.00 | 71.52 |
| yi-34b-hf | 89.81 | 86.55 | 83.92 | 57.23 | 64.04 | 73.64 | 79.49 | 85.50 | 66.40 | 52.32 | 81.00 | 86.06 |
| deepseek-7b-base-hf | 55.56 | 73.10 | 56.59 | 46.99 | 34.98 | 62.73 | 48.21 | 58.78 | 28.57 | 29.14 | 50.00 | 61.82 |
| deepseek-67b-base-hf | 84.26 | 85.96 | 81.03 | 56.02 | 57.64 | 72.73 | 73.85 | 82.44 | 51.59 | 45.03 | 74.00 | 81.82 |
| model | business_ethics | moral_disputes | high_school_statistics | miscellaneous | formal_logic | high_school_government_and_politics | prehistory | security_studies | high_school_biology | logical_fallacies | high_school_world_history | professional_medicine |
|:------------------------:|------------------:|-----------------:|-------------------------:|----------------:|---------------:|--------------------------------------:|-------------:|-------------------:|----------------------:|--------------------:|----------------------------:|------------------------:|
| llama-7b-turbomind | 42.00 | 40.46 | 32.87 | 42.78 | 26.19 | 46.11 | 35.19 | 33.47 | 32.90 | 42.33 | 43.88 | 43.75 |
| llama-13b-turbomind | 46.00 | 50.00 | 30.56 | 64.88 | 31.75 | 66.84 | 51.85 | 52.65 | 51.94 | 52.76 | 67.51 | 51.10 |
| llama-30b-turbomind | 55.00 | 66.76 | 49.07 | 77.91 | 36.51 | 82.90 | 68.21 | 66.12 | 69.35 | 67.48 | 80.59 | 55.88 |
| llama-65b-turbomind | 59.00 | 73.70 | 61.57 | 81.35 | 43.65 | 88.60 | 73.46 | 71.84 | 74.19 | 77.30 | 83.97 | 62.13 |
| llama-2-7b-turbomind | 53.00 | 51.16 | 27.78 | 63.60 | 27.78 | 67.36 | 48.77 | 47.76 | 50.97 | 51.53 | 64.56 | 52.57 |
| llama-2-13b-turbomind | 54.00 | 64.45 | 45.37 | 74.46 | 36.51 | 80.83 | 64.81 | 62.86 | 67.42 | 66.87 | 72.15 | 54.41 |
| llama-2-70b-turbomind | 72.00 | 77.17 | 63.43 | 86.08 | 48.41 | 94.30 | 83.64 | 78.37 | 81.61 | 80.98 | 87.76 | 74.63 |
| llama-3-8b-turbomind | 62.00 | 73.70 | 54.17 | 82.76 | 48.41 | 90.16 | 72.53 | 75.51 | 77.74 | 73.01 | 82.70 | 72.06 |
| llama-3-70b-turbomind | 83.00 | 85.55 | 72.22 | 92.21 | 66.67 | 97.41 | 91.05 | 84.90 | 90.32 | 87.73 | 94.09 | 87.13 |
| internlm2-1.8b-turbomind | 44.00 | 45.95 | 38.89 | 59.39 | 32.54 | 60.62 | 50.31 | 54.29 | 52.58 | 45.40 | 62.87 | 37.87 |
| internlm2-7b-turbomind | 69.00 | 66.76 | 57.87 | 80.72 | 50.00 | 90.16 | 73.15 | 75.10 | 79.68 | 68.71 | 81.01 | 70.22 |
| internlm2-20b-turbomind | 74.00 | 74.57 | 60.19 | 81.48 | 44.44 | 91.71 | 75.31 | 81.63 | 82.58 | 75.46 | 87.76 | 63.60 |
| qwen-1.8b-turbomind | 52.00 | 52.31 | 34.72 | 57.98 | 29.37 | 59.07 | 47.22 | 48.57 | 52.26 | 44.17 | 61.18 | 43.38 |
| qwen-7b-turbomind | 68.00 | 64.74 | 45.37 | 77.39 | 43.65 | 83.94 | 68.21 | 70.20 | 72.26 | 65.64 | 75.95 | 58.46 |
| qwen-14b-turbomind | 75.00 | 74.86 | 57.87 | 84.04 | 51.59 | 91.71 | 70.99 | 77.14 | 83.55 | 73.01 | 83.12 | 67.65 |
| qwen-72b-turbomind | 80.00 | 84.97 | 68.98 | 91.44 | 54.76 | 98.96 | 87.04 | 81.63 | 89.03 | 84.05 | 90.30 | 84.93 |
| qwen1.5-0.5b-hf | 47.00 | 46.82 | 23.15 | 48.02 | 29.37 | 48.70 | 40.12 | 38.37 | 40.65 | 35.58 | 53.16 | 31.62 |
| qwen1.5-1.8b-hf | 54.00 | 54.91 | 28.70 | 61.69 | 23.81 | 58.03 | 48.15 | 51.84 | 55.48 | 45.40 | 59.92 | 39.71 |
| qwen1.5-4b-hf | 65.00 | 66.76 | 44.44 | 73.95 | 35.71 | 78.24 | 60.19 | 65.31 | 66.45 | 65.64 | 71.31 | 50.00 |
| qwen1.5-7b-hf | 68.00 | 70.81 | 48.61 | 76.50 | 38.89 | 84.97 | 69.44 | 68.16 | 74.52 | 68.10 | 77.22 | 56.25 |
| qwen1.5-14b-hf | 77.00 | 73.70 | 62.96 | 83.40 | 53.17 | 90.67 | 71.60 | 80.82 | 84.52 | 76.69 | 83.54 | 71.69 |
| qwen1.5-32b-hf | 77.00 | 78.90 | 68.98 | 88.12 | 54.76 | 94.82 | 81.48 | 80.82 | 88.39 | 82.21 | 86.08 | 80.88 |
| qwen1.5-72b-hf | 80.00 | 84.39 | 68.98 | 91.44 | 55.56 | 98.96 | 86.73 | 81.63 | 88.71 | 85.89 | 89.87 | 82.72 |
| qwen1.5-moe-a2-7b-hf | 74.00 | 65.90 | 56.48 | 82.25 | 34.13 | 84.46 | 70.68 | 74.29 | 73.23 | 68.10 | 76.79 | 66.91 |
| mistral-7b-v0.1-hf | 57.00 | 71.10 | 57.41 | 81.61 | 40.48 | 86.53 | 73.46 | 72.65 | 76.77 | 79.14 | 77.22 | 68.75 |
| mistral-7b-v0.2-hf | 61.00 | 71.39 | 52.78 | 80.08 | 40.48 | 88.08 | 69.44 | 72.24 | 76.13 | 77.91 | 78.06 | 70.59 |
| mixtral-8x7b-v0.1-hf | 77.00 | 80.06 | 63.43 | 87.87 | 54.76 | 93.26 | 83.95 | 80.00 | 84.19 | 79.14 | 88.61 | 81.25 |
| mixtral-8x22b-v0.1-hf | 72.00 | 84.10 | 68.52 | 90.68 | 57.14 | 96.37 | 86.73 | 86.53 | 90.32 | 87.73 | 90.30 | 87.87 |
| yi-6b-hf | 67.00 | 69.36 | 52.78 | 80.46 | 44.44 | 89.64 | 70.99 | 74.69 | 77.10 | 78.53 | 78.90 | 65.81 |
| yi-34b-hf | 79.00 | 83.82 | 66.67 | 90.29 | 57.14 | 97.93 | 87.65 | 84.90 | 88.39 | 87.73 | 92.83 | 81.99 |
| deepseek-7b-base-hf | 49.00 | 52.31 | 41.20 | 66.28 | 30.95 | 63.73 | 55.86 | 51.84 | 52.90 | 58.90 | 62.45 | 45.22 |
| deepseek-67b-base-hf | 81.00 | 77.17 | 63.89 | 90.04 | 53.17 | 97.93 | 85.49 | 73.88 | 82.26 | 84.05 | 91.56 | 78.31 |
| model | high_school_mathematics | college_medicine | high_school_us_history | sociology | econometrics | high_school_psychology | human_aging | us_foreign_policy |
|:------------------------:|--------------------------:|-------------------:|-------------------------:|------------:|---------------:|-------------------------:|--------------:|--------------------:|
| llama-7b-turbomind | 24.81 | 32.95 | 38.73 | 45.77 | 27.19 | 48.07 | 38.12 | 43.00 |
| llama-13b-turbomind | 26.30 | 42.20 | 59.80 | 61.19 | 28.95 | 61.28 | 53.36 | 78.00 |
| llama-30b-turbomind | 27.41 | 54.91 | 76.96 | 79.10 | 35.96 | 76.15 | 67.71 | 83.00 |
| llama-65b-turbomind | 34.44 | 54.34 | 82.84 | 81.09 | 39.47 | 82.39 | 66.37 | 88.00 |
| llama-2-7b-turbomind | 29.63 | 43.35 | 60.29 | 62.69 | 27.19 | 62.75 | 56.05 | 64.00 |
| llama-2-13b-turbomind | 27.04 | 52.60 | 75.49 | 73.13 | 32.46 | 76.51 | 64.57 | 82.00 |
| llama-2-70b-turbomind | 34.07 | 64.16 | 90.69 | 90.55 | 44.74 | 87.52 | 80.27 | 92.00 |
| llama-3-8b-turbomind | 38.15 | 64.16 | 83.33 | 86.57 | 47.37 | 84.04 | 70.85 | 87.00 |
| llama-3-70b-turbomind | 48.89 | 79.77 | 95.10 | 94.03 | 72.81 | 94.13 | 82.51 | 94.00 |
| internlm2-1.8b-turbomind | 30.37 | 41.04 | 55.88 | 51.74 | 28.95 | 61.47 | 51.12 | 63.00 |
| internlm2-7b-turbomind | 39.63 | 68.21 | 76.96 | 84.58 | 44.74 | 84.59 | 72.65 | 86.00 |
| internlm2-20b-turbomind | 39.63 | 66.47 | 82.84 | 85.07 | 47.37 | 86.79 | 70.85 | 84.00 |
| qwen-1.8b-turbomind | 28.52 | 43.35 | 54.90 | 60.70 | 36.84 | 60.73 | 48.43 | 60.00 |
| qwen-7b-turbomind | 30.00 | 57.23 | 75.98 | 79.10 | 32.46 | 79.27 | 63.23 | 81.00 |
| qwen-14b-turbomind | 37.41 | 70.52 | 81.37 | 85.07 | 50.00 | 84.95 | 73.09 | 86.00 |
| qwen-72b-turbomind | 50.00 | 75.72 | 92.16 | 90.05 | 59.65 | 92.66 | 82.51 | 95.00 |
| qwen1.5-0.5b-hf | 29.63 | 33.53 | 45.10 | 59.70 | 28.95 | 44.77 | 37.22 | 69.00 |
| qwen1.5-1.8b-hf | 34.07 | 39.31 | 47.55 | 63.18 | 32.46 | 59.08 | 53.81 | 73.00 |
| qwen1.5-4b-hf | 35.93 | 55.49 | 71.08 | 73.13 | 37.72 | 72.11 | 63.68 | 79.00 |
| qwen1.5-7b-hf | 34.81 | 61.85 | 78.92 | 82.09 | 41.23 | 80.73 | 61.88 | 84.00 |
| qwen1.5-14b-hf | 45.93 | 68.21 | 80.88 | 83.08 | 55.26 | 86.06 | 73.09 | 88.00 |
| qwen1.5-32b-hf | 47.04 | 76.30 | 90.20 | 86.07 | 57.89 | 90.28 | 75.78 | 92.00 |
| qwen1.5-72b-hf | 47.78 | 75.14 | 92.65 | 88.56 | 59.65 | 92.48 | 79.82 | 94.00 |
| qwen1.5-moe-a2-7b-hf | 46.30 | 54.91 | 78.43 | 79.10 | 38.60 | 82.39 | 66.82 | 83.00 |
| mistral-7b-v0.1-hf | 33.70 | 65.32 | 78.92 | 83.08 | 50.00 | 82.39 | 69.51 | 86.00 |
| mistral-7b-v0.2-hf | 38.15 | 64.16 | 81.86 | 82.09 | 43.86 | 80.18 | 69.96 | 86.00 |
| mixtral-8x7b-v0.1-hf | 40.37 | 69.94 | 86.27 | 88.56 | 65.79 | 88.81 | 79.37 | 91.00 |
| mixtral-8x22b-v0.1-hf | 45.93 | 79.19 | 90.20 | 93.03 | 70.18 | 92.29 | 79.37 | 95.00 |
| yi-6b-hf | 32.59 | 61.27 | 79.90 | 82.59 | 35.96 | 82.94 | 67.26 | 86.00 |
| yi-34b-hf | 45.19 | 71.68 | 91.18 | 88.56 | 55.26 | 91.74 | 78.48 | 91.00 |
| deepseek-7b-base-hf | 28.89 | 41.62 | 60.29 | 70.15 | 26.32 | 69.72 | 55.61 | 76.00 |
| deepseek-67b-base-hf | 38.89 | 72.25 | 90.69 | 90.05 | 52.63 | 90.46 | 80.72 | 95.00 |
## Chat Models
| model | mmlu | mmlu-stem | mmlu-social-science | mmlu-humanities | mmlu-other |
|:-----------------------------:|-------:|------------:|----------------------:|------------------:|-------------:|
| qwen1.5-0.5b-chat-hf | 35.32 | 30.90 | 37.59 | 37.29 | 37.73 |
| qwen1.5-1.8b-chat-hf | 45.62 | 39.20 | 49.21 | 47.67 | 49.63 |
| qwen1.5-4b-chat-hf | 55.90 | 48.07 | 62.67 | 59.70 | 57.31 |
| qwen1.5-7b-chat-hf | 61.79 | 52.68 | 69.41 | 66.41 | 63.45 |
| qwen1.5-14b-chat-hf | 67.96 | 59.79 | 75.46 | 71.23 | 69.72 |
| qwen1.5-32b-chat-hf | 75.36 | 67.04 | 82.11 | 80.44 | 76.23 |
| qwen1.5-72b-chat-hf | 77.24 | 69.59 | 83.95 | 81.58 | 77.87 |
| qwen1.5-110b-chat-hf | 77.95 | 71.56 | 83.77 | 81.44 | 78.41 |
| internlm2-chat-1.8b-hf | 47.58 | 40.88 | 53.33 | 49.92 | 49.74 |
| internlm2-chat-1.8b-sft-hf | 47.44 | 40.55 | 53.31 | 49.67 | 49.89 |
| internlm2-chat-7b-hf | 63.05 | 53.42 | 71.47 | 67.27 | 65.13 |
| internlm2-chat-7b-sft-hf | 63.33 | 53.95 | 71.74 | 67.62 | 65.00 |
| internlm2-chat-20b-hf | 67.37 | 57.39 | 75.75 | 71.63 | 69.95 |
| internlm2-chat-20b-sft-hf | 67.34 | 57.49 | 75.67 | 70.99 | 70.40 |
| llama-3-8b-instruct-hf | 68.37 | 58.01 | 77.82 | 71.22 | 71.94 |
| llama-3-70b-instruct-hf | 80.93 | 73.86 | 87.71 | 83.90 | 82.01 |
| llama-3-8b-instruct-lmdeploy | 67.35 | 56.66 | 75.96 | 70.90 | 71.49 |
| llama-3-70b-instruct-lmdeploy | 80.85 | 74.07 | 87.26 | 83.73 | 81.96 |
| mistral-7b-instruct-v0.1-hf | 54.36 | 43.74 | 62.96 | 58.87 | 57.46 |
| mistral-7b-instruct-v0.2-hf | 59.98 | 49.56 | 69.22 | 64.41 | 62.24 |
| mixtral-8x7b-instruct-v0.1-hf | 70.11 | 60.29 | 79.01 | 74.08 | 72.28 |
### Details
| model | college_biology | college_chemistry | college_computer_science | college_mathematics | college_physics | electrical_engineering | astronomy | anatomy | abstract_algebra | machine_learning | clinical_knowledge | global_facts |
|:-----------------------------:|------------------:|--------------------:|---------------------------:|----------------------:|------------------:|-------------------------:|------------:|----------:|-------------------:|-------------------:|---------------------:|---------------:|
| qwen1.5-0.5b-chat-hf | 31.25 | 32.00 | 33.00 | 29.00 | 33.33 | 38.62 | 33.55 | 28.89 | 20.00 | 27.68 | 40.38 | 33.00 |
| qwen1.5-1.8b-chat-hf | 42.36 | 28.00 | 45.00 | 33.00 | 27.45 | 44.83 | 51.97 | 42.22 | 32.00 | 38.39 | 48.30 | 30.00 |
| qwen1.5-4b-chat-hf | 56.25 | 47.00 | 49.00 | 39.00 | 36.27 | 54.48 | 57.89 | 49.63 | 38.00 | 33.04 | 59.62 | 23.00 |
| qwen1.5-7b-chat-hf | 64.58 | 51.00 | 59.00 | 37.00 | 41.18 | 53.79 | 66.45 | 53.33 | 43.00 | 41.07 | 67.92 | 36.00 |
| qwen1.5-14b-chat-hf | 77.08 | 51.00 | 64.00 | 42.00 | 45.10 | 64.83 | 77.63 | 65.93 | 39.00 | 46.43 | 73.21 | 45.00 |
| qwen1.5-32b-chat-hf | 84.72 | 53.00 | 57.00 | 48.00 | 52.94 | 74.48 | 82.24 | 67.41 | 52.00 | 61.61 | 78.11 | 48.00 |
| qwen1.5-72b-chat-hf | 90.97 | 57.00 | 66.00 | 55.00 | 55.88 | 80.00 | 88.16 | 72.59 | 56.00 | 59.82 | 80.00 | 51.00 |
| qwen1.5-110b-chat-hf | 88.89 | 62.00 | 66.00 | 64.00 | 58.82 | 75.86 | 89.47 | 68.15 | 59.00 | 63.39 | 79.62 | 59.00 |
| internlm2-chat-1.8b-hf | 49.31 | 36.00 | 47.00 | 33.00 | 36.27 | 42.76 | 48.03 | 49.63 | 30.00 | 33.93 | 53.58 | 28.00 |
| internlm2-chat-1.8b-sft-hf | 51.39 | 37.00 | 50.00 | 33.00 | 33.33 | 42.76 | 46.05 | 49.63 | 31.00 | 32.14 | 53.21 | 29.00 |
| internlm2-chat-7b-hf | 68.75 | 47.00 | 62.00 | 32.00 | 38.24 | 57.24 | 69.74 | 58.52 | 29.00 | 53.57 | 70.19 | 41.00 |
| internlm2-chat-7b-sft-hf | 71.53 | 47.00 | 63.00 | 34.00 | 37.25 | 57.24 | 69.74 | 57.78 | 29.00 | 52.68 | 69.43 | 34.00 |
| internlm2-chat-20b-hf | 76.39 | 51.00 | 61.00 | 37.00 | 40.20 | 62.76 | 78.95 | 67.41 | 33.00 | 46.43 | 75.09 | 42.00 |
| internlm2-chat-20b-sft-hf | 77.08 | 49.00 | 60.00 | 39.00 | 39.22 | 64.14 | 79.61 | 68.15 | 35.00 | 46.43 | 75.09 | 42.00 |
| llama-3-8b-instruct-hf | 81.94 | 48.00 | 58.00 | 43.00 | 48.04 | 60.69 | 76.32 | 71.11 | 33.00 | 54.46 | 73.58 | 46.00 |
| llama-3-70b-instruct-hf | 93.06 | 56.00 | 70.00 | 60.00 | 60.78 | 77.24 | 93.42 | 79.26 | 53.00 | 71.43 | 86.42 | 66.00 |
| llama-3-8b-instruct-lmdeploy | 79.17 | 47.00 | 53.00 | 36.00 | 49.02 | 60.00 | 73.68 | 68.89 | 36.00 | 55.36 | 73.96 | 42.00 |
| llama-3-70b-instruct-lmdeploy | 93.75 | 57.00 | 66.00 | 61.00 | 65.69 | 77.93 | 92.11 | 78.52 | 55.00 | 70.54 | 86.42 | 64.00 |
| mistral-7b-instruct-v0.1-hf | 57.64 | 35.00 | 50.00 | 31.00 | 24.51 | 51.72 | 58.55 | 45.93 | 35.00 | 41.07 | 56.98 | 32.00 |
| mistral-7b-instruct-v0.2-hf | 70.14 | 42.00 | 49.00 | 35.00 | 43.14 | 54.48 | 65.79 | 56.30 | 29.00 | 42.86 | 65.28 | 37.00 |
| mixtral-8x7b-instruct-v0.1-hf | 81.25 | 57.00 | 57.00 | 40.00 | 50.00 | 60.69 | 80.92 | 65.93 | 45.00 | 50.89 | 76.60 | 41.00 |
| model | management | nutrition | marketing | professional_accounting | high_school_geography | international_law | moral_scenarios | computer_security | high_school_microeconomics | professional_law | medical_genetics | professional_psychology |
|:-----------------------------:|-------------:|------------:|------------:|--------------------------:|------------------------:|--------------------:|------------------:|--------------------:|-----------------------------:|-------------------:|-------------------:|--------------------------:|
| qwen1.5-0.5b-chat-hf | 41.75 | 38.89 | 49.15 | 26.60 | 48.48 | 50.41 | 24.69 | 42.00 | 32.35 | 31.75 | 31.00 | 32.35 |
| qwen1.5-1.8b-chat-hf | 62.14 | 55.56 | 76.92 | 34.40 | 58.08 | 61.16 | 21.90 | 56.00 | 42.44 | 35.14 | 50.00 | 44.93 |
| qwen1.5-4b-chat-hf | 73.79 | 58.50 | 82.05 | 47.16 | 74.24 | 71.90 | 32.29 | 69.00 | 58.40 | 40.74 | 58.00 | 53.76 |
| qwen1.5-7b-chat-hf | 79.61 | 69.28 | 85.47 | 41.49 | 78.79 | 76.86 | 35.75 | 74.00 | 65.13 | 44.78 | 68.00 | 57.68 |
| qwen1.5-14b-chat-hf | 82.52 | 70.26 | 87.18 | 51.77 | 85.86 | 82.64 | 53.74 | 81.00 | 76.05 | 47.98 | 76.00 | 67.48 |
| qwen1.5-32b-chat-hf | 84.47 | 77.78 | 94.44 | 60.99 | 90.91 | 87.60 | 72.96 | 79.00 | 83.61 | 58.28 | 83.00 | 77.94 |
| qwen1.5-72b-chat-hf | 89.32 | 85.95 | 93.59 | 61.35 | 90.91 | 86.78 | 75.98 | 83.00 | 84.87 | 60.30 | 83.00 | 81.05 |
| qwen1.5-110b-chat-hf | 86.41 | 80.72 | 92.74 | 69.15 | 93.94 | 84.30 | 77.88 | 83.00 | 88.66 | 61.73 | 84.00 | 82.19 |
| internlm2-chat-1.8b-hf | 72.82 | 50.65 | 69.23 | 35.46 | 56.06 | 56.20 | 27.82 | 60.00 | 49.16 | 33.83 | 54.00 | 43.79 |
| internlm2-chat-1.8b-sft-hf | 71.84 | 52.61 | 68.80 | 34.75 | 55.56 | 53.72 | 27.04 | 58.00 | 48.74 | 34.09 | 54.00 | 44.61 |
| internlm2-chat-7b-hf | 78.64 | 66.67 | 85.90 | 46.81 | 79.29 | 70.25 | 35.31 | 79.00 | 68.07 | 46.41 | 68.00 | 64.87 |
| internlm2-chat-7b-sft-hf | 79.61 | 67.97 | 86.75 | 47.52 | 80.30 | 70.25 | 35.98 | 80.00 | 69.33 | 45.83 | 70.00 | 65.36 |
| internlm2-chat-20b-hf | 80.58 | 75.16 | 90.17 | 52.13 | 83.84 | 80.99 | 39.33 | 80.00 | 70.59 | 49.67 | 75.00 | 70.26 |
| internlm2-chat-20b-sft-hf | 80.58 | 76.14 | 91.03 | 53.19 | 84.34 | 80.99 | 36.31 | 77.00 | 71.85 | 49.61 | 77.00 | 70.59 |
| llama-3-8b-instruct-hf | 82.52 | 79.41 | 91.45 | 52.48 | 80.30 | 79.34 | 46.26 | 75.00 | 76.89 | 49.61 | 85.00 | 72.22 |
| llama-3-70b-instruct-hf | 89.32 | 87.58 | 93.16 | 66.67 | 92.42 | 90.08 | 76.20 | 83.00 | 89.50 | 64.67 | 92.00 | 87.09 |
| llama-3-8b-instruct-lmdeploy | 87.38 | 79.41 | 90.17 | 52.48 | 79.80 | 78.51 | 44.25 | 75.00 | 74.37 | 48.76 | 84.00 | 69.61 |
| llama-3-70b-instruct-lmdeploy | 90.29 | 88.56 | 93.59 | 65.96 | 92.93 | 89.26 | 75.75 | 83.00 | 89.92 | 63.95 | 92.00 | 86.60 |
| mistral-7b-instruct-v0.1-hf | 69.90 | 59.80 | 85.47 | 38.65 | 69.70 | 65.29 | 37.54 | 69.00 | 51.26 | 37.81 | 65.00 | 52.45 |
| mistral-7b-instruct-v0.2-hf | 74.76 | 66.99 | 88.89 | 43.97 | 75.25 | 76.86 | 42.01 | 73.00 | 62.61 | 42.24 | 67.00 | 62.25 |
| mixtral-8x7b-instruct-v0.1-hf | 85.44 | 80.39 | 92.74 | 55.32 | 85.35 | 82.64 | 48.38 | 78.00 | 75.21 | 53.52 | 75.00 | 74.02 |
| model | jurisprudence | world_religions | philosophy | virology | high_school_chemistry | public_relations | high_school_macroeconomics | human_sexuality | elementary_mathematics | high_school_physics | high_school_computer_science | high_school_european_history |
|:-----------------------------:|----------------:|------------------:|-------------:|-----------:|------------------------:|-------------------:|-----------------------------:|------------------:|-------------------------:|----------------------:|-------------------------------:|-------------------------------:|
| qwen1.5-0.5b-chat-hf | 42.59 | 24.56 | 39.87 | 39.76 | 29.06 | 38.18 | 35.64 | 38.93 | 27.78 | 29.80 | 34.00 | 48.48 |
| qwen1.5-1.8b-chat-hf | 50.93 | 56.73 | 44.37 | 42.77 | 35.96 | 51.82 | 38.46 | 49.62 | 35.45 | 27.15 | 47.00 | 63.03 |
| qwen1.5-4b-chat-hf | 71.30 | 65.50 | 58.20 | 50.00 | 44.33 | 57.27 | 54.10 | 61.83 | 43.65 | 41.06 | 60.00 | 72.12 |
| qwen1.5-7b-chat-hf | 76.85 | 76.61 | 68.49 | 48.80 | 51.72 | 64.55 | 59.23 | 68.70 | 48.94 | 37.09 | 69.00 | 79.39 |
| qwen1.5-14b-chat-hf | 75.93 | 80.70 | 69.13 | 51.20 | 55.67 | 64.55 | 67.69 | 74.05 | 57.14 | 47.02 | 74.00 | 82.42 |
| qwen1.5-32b-chat-hf | 83.33 | 89.47 | 82.64 | 60.84 | 62.56 | 70.00 | 76.67 | 83.21 | 67.46 | 59.60 | 85.00 | 84.85 |
| qwen1.5-72b-chat-hf | 86.11 | 89.47 | 80.71 | 59.04 | 68.47 | 72.73 | 80.00 | 87.79 | 67.72 | 52.32 | 79.00 | 85.45 |
| qwen1.5-110b-chat-hf | 83.33 | 87.13 | 81.03 | 54.22 | 69.95 | 73.64 | 78.21 | 87.02 | 75.93 | 57.62 | 84.00 | 88.48 |
| internlm2-chat-1.8b-hf | 52.78 | 60.82 | 49.20 | 42.77 | 42.36 | 50.00 | 47.18 | 53.44 | 32.54 | 31.79 | 39.00 | 60.00 |
| internlm2-chat-1.8b-sft-hf | 53.70 | 61.40 | 50.16 | 42.17 | 40.89 | 50.00 | 47.69 | 51.15 | 32.54 | 29.14 | 40.00 | 59.39 |
| internlm2-chat-7b-hf | 73.15 | 81.87 | 67.85 | 47.59 | 49.75 | 62.73 | 61.79 | 66.41 | 44.97 | 33.77 | 71.00 | 81.82 |
| internlm2-chat-7b-sft-hf | 73.15 | 81.87 | 66.88 | 48.19 | 48.77 | 63.64 | 62.31 | 65.65 | 45.77 | 33.77 | 72.00 | 81.82 |
| internlm2-chat-20b-hf | 80.56 | 81.87 | 72.99 | 55.42 | 54.19 | 70.00 | 67.95 | 71.76 | 48.15 | 39.74 | 75.00 | 80.00 |
| internlm2-chat-20b-sft-hf | 81.48 | 79.53 | 72.99 | 54.82 | 54.19 | 69.09 | 67.95 | 71.76 | 48.94 | 41.06 | 75.00 | 80.00 |
| llama-3-8b-instruct-hf | 76.85 | 79.53 | 72.35 | 53.61 | 54.19 | 70.91 | 66.41 | 80.92 | 49.47 | 46.36 | 71.00 | 75.15 |
| llama-3-70b-instruct-hf | 87.04 | 88.30 | 82.64 | 56.02 | 67.49 | 74.55 | 86.41 | 88.55 | 74.34 | 65.56 | 91.00 | 86.06 |
| llama-3-8b-instruct-lmdeploy | 77.78 | 79.53 | 70.74 | 52.41 | 53.20 | 68.18 | 65.38 | 79.39 | 50.79 | 37.75 | 72.00 | 76.97 |
| llama-3-70b-instruct-lmdeploy | 87.96 | 90.64 | 83.28 | 54.82 | 69.46 | 73.64 | 86.92 | 87.02 | 74.87 | 66.23 | 92.00 | 85.45 |
| mistral-7b-instruct-v0.1-hf | 64.81 | 70.18 | 63.67 | 41.57 | 38.92 | 68.18 | 49.49 | 61.83 | 33.33 | 32.45 | 55.00 | 66.67 |
| mistral-7b-instruct-v0.2-hf | 70.37 | 80.12 | 64.95 | 50.60 | 50.74 | 68.18 | 54.36 | 71.76 | 40.74 | 35.10 | 60.00 | 73.33 |
| mixtral-8x7b-instruct-v0.1-hf | 79.63 | 87.72 | 73.63 | 54.82 | 61.58 | 67.27 | 69.49 | 83.21 | 52.91 | 47.02 | 74.00 | 80.61 |
| model | business_ethics | moral_disputes | high_school_statistics | miscellaneous | formal_logic | high_school_government_and_politics | prehistory | security_studies | high_school_biology | logical_fallacies | high_school_world_history | professional_medicine |
|:-----------------------------:|------------------:|-----------------:|-------------------------:|----------------:|---------------:|--------------------------------------:|-------------:|-------------------:|----------------------:|--------------------:|----------------------------:|------------------------:|
| qwen1.5-0.5b-chat-hf | 45.00 | 41.04 | 30.09 | 39.21 | 24.60 | 35.23 | 33.95 | 25.31 | 36.13 | 31.29 | 49.37 | 38.24 |
| qwen1.5-1.8b-chat-hf | 54.00 | 50.29 | 34.26 | 58.49 | 24.60 | 55.96 | 47.53 | 39.18 | 47.74 | 44.17 | 64.98 | 40.81 |
| qwen1.5-4b-chat-hf | 61.00 | 64.16 | 46.30 | 71.01 | 39.68 | 72.02 | 54.01 | 65.31 | 63.55 | 63.80 | 71.31 | 51.10 |
| qwen1.5-7b-chat-hf | 69.00 | 67.05 | 50.93 | 76.25 | 53.17 | 82.38 | 62.96 | 71.02 | 73.23 | 68.10 | 76.79 | 60.29 |
| qwen1.5-14b-chat-hf | 74.00 | 75.14 | 58.33 | 82.89 | 51.59 | 88.60 | 69.44 | 77.96 | 84.19 | 73.62 | 82.70 | 71.32 |
| qwen1.5-32b-chat-hf | 80.00 | 80.64 | 70.83 | 89.40 | 60.32 | 94.82 | 81.79 | 79.59 | 90.00 | 86.50 | 88.61 | 80.15 |
| qwen1.5-72b-chat-hf | 80.00 | 82.95 | 68.98 | 91.83 | 57.14 | 98.45 | 86.73 | 78.78 | 89.03 | 87.12 | 91.14 | 83.82 |
| qwen1.5-110b-chat-hf | 79.00 | 78.03 | 67.13 | 92.98 | 62.70 | 97.93 | 87.04 | 74.29 | 88.71 | 82.82 | 91.14 | 84.93 |
| internlm2-chat-1.8b-hf | 48.00 | 49.13 | 44.91 | 57.60 | 26.98 | 61.14 | 50.62 | 51.02 | 52.58 | 57.67 | 67.51 | 37.50 |
| internlm2-chat-1.8b-sft-hf | 50.00 | 49.13 | 44.91 | 57.73 | 28.57 | 61.66 | 49.69 | 51.02 | 49.68 | 57.67 | 66.67 | 38.60 |
| internlm2-chat-7b-hf | 65.00 | 65.61 | 49.54 | 80.84 | 43.65 | 88.08 | 70.99 | 68.98 | 78.39 | 75.46 | 82.28 | 61.76 |
| internlm2-chat-7b-sft-hf | 64.00 | 66.18 | 52.31 | 81.35 | 46.03 | 88.08 | 71.60 | 67.76 | 78.39 | 77.30 | 82.28 | 63.60 |
| internlm2-chat-20b-hf | 74.00 | 73.70 | 59.72 | 81.86 | 46.83 | 89.12 | 74.69 | 75.92 | 80.65 | 79.14 | 82.70 | 70.59 |
| internlm2-chat-20b-sft-hf | 76.00 | 73.12 | 60.19 | 81.99 | 43.65 | 88.60 | 74.38 | 73.88 | 80.32 | 80.37 | 82.70 | 70.59 |
| llama-3-8b-instruct-hf | 72.00 | 73.12 | 55.09 | 84.55 | 50.00 | 90.67 | 77.16 | 77.55 | 81.61 | 77.91 | 84.81 | 75.00 |
| llama-3-70b-instruct-hf | 85.00 | 85.26 | 75.00 | 92.72 | 69.05 | 97.41 | 90.43 | 82.04 | 91.61 | 87.12 | 94.09 | 89.71 |
| llama-3-8b-instruct-lmdeploy | 72.00 | 72.83 | 52.78 | 82.12 | 51.59 | 89.64 | 76.85 | 76.73 | 80.97 | 76.69 | 84.39 | 74.63 |
| llama-3-70b-instruct-lmdeploy | 85.00 | 84.39 | 73.61 | 92.72 | 67.46 | 97.93 | 89.81 | 81.63 | 90.65 | 87.12 | 93.25 | 89.34 |
| mistral-7b-instruct-v0.1-hf | 55.00 | 57.51 | 39.81 | 74.07 | 39.68 | 75.65 | 57.72 | 62.04 | 59.35 | 69.33 | 67.93 | 55.88 |
| mistral-7b-instruct-v0.2-hf | 61.00 | 66.76 | 46.76 | 78.67 | 36.51 | 84.97 | 68.83 | 70.20 | 68.39 | 69.33 | 73.00 | 58.09 |
| mixtral-8x7b-instruct-v0.1-hf | 66.00 | 76.59 | 57.87 | 86.59 | 50.00 | 93.78 | 83.02 | 79.18 | 82.58 | 75.46 | 86.50 | 77.94 |
| model | high_school_mathematics | college_medicine | high_school_us_history | sociology | econometrics | high_school_psychology | human_aging | us_foreign_policy |
|:-----------------------------:|--------------------------:|-------------------:|-------------------------:|------------:|---------------:|-------------------------:|--------------:|--------------------:|
| qwen1.5-0.5b-chat-hf | 24.44 | 35.26 | 42.16 | 47.26 | 29.82 | 40.55 | 32.29 | 47.00 |
| qwen1.5-1.8b-chat-hf | 32.22 | 43.35 | 54.90 | 48.26 | 28.95 | 61.83 | 48.43 | 71.00 |
| qwen1.5-4b-chat-hf | 36.30 | 51.45 | 71.08 | 76.62 | 34.21 | 72.29 | 58.30 | 72.00 |
| qwen1.5-7b-chat-hf | 31.11 | 61.27 | 76.47 | 79.10 | 42.11 | 81.28 | 61.43 | 83.00 |
| qwen1.5-14b-chat-hf | 41.48 | 68.79 | 80.88 | 82.59 | 48.25 | 84.40 | 72.20 | 88.00 |
| qwen1.5-32b-chat-hf | 48.52 | 75.72 | 88.73 | 86.07 | 57.02 | 90.46 | 78.03 | 95.00 |
| qwen1.5-72b-chat-hf | 51.48 | 73.99 | 90.69 | 87.06 | 59.65 | 92.11 | 79.37 | 94.00 |
| qwen1.5-110b-chat-hf | 52.22 | 76.30 | 93.14 | 87.56 | 62.28 | 91.56 | 80.27 | 88.00 |
| internlm2-chat-1.8b-hf | 31.48 | 46.82 | 56.37 | 65.17 | 28.07 | 65.87 | 50.22 | 69.00 |
| internlm2-chat-1.8b-sft-hf | 30.74 | 47.40 | 54.41 | 64.18 | 29.82 | 66.24 | 48.43 | 69.00 |
| internlm2-chat-7b-hf | 33.70 | 67.05 | 79.90 | 81.09 | 48.25 | 84.04 | 67.26 | 84.00 |
| internlm2-chat-7b-sft-hf | 35.19 | 67.05 | 79.90 | 80.60 | 48.25 | 84.59 | 65.47 | 85.00 |
| internlm2-chat-20b-hf | 36.30 | 66.47 | 88.73 | 85.07 | 51.75 | 85.69 | 70.85 | 87.00 |
| internlm2-chat-20b-sft-hf | 35.93 | 65.90 | 87.75 | 85.57 | 52.63 | 84.77 | 70.85 | 87.00 |
| llama-3-8b-instruct-hf | 36.67 | 68.79 | 83.82 | 86.57 | 61.40 | 84.95 | 70.85 | 85.00 |
| llama-3-70b-instruct-hf | 57.41 | 78.61 | 89.71 | 91.54 | 74.56 | 94.50 | 82.96 | 94.00 |
| llama-3-8b-instruct-lmdeploy | 38.52 | 68.79 | 82.84 | 85.57 | 54.39 | 85.50 | 69.96 | 83.00 |
| llama-3-70b-instruct-lmdeploy | 54.81 | 79.77 | 90.20 | 92.04 | 71.05 | 94.50 | 82.96 | 93.00 |
| mistral-7b-instruct-v0.1-hf | 28.89 | 50.29 | 67.16 | 76.12 | 39.47 | 72.29 | 62.33 | 77.00 |
| mistral-7b-instruct-v0.2-hf | 30.74 | 53.18 | 73.04 | 77.11 | 42.11 | 79.82 | 63.68 | 82.00 |
| mixtral-8x7b-instruct-v0.1-hf | 35.56 | 73.41 | 85.29 | 87.06 | 60.53 | 86.97 | 74.44 | 86.00 |

View File

@ -0,0 +1,69 @@
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import FixKRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.openicl.icl_evaluator import AccwithDetailsEvaluator
from opencompass.datasets import MMLUDataset
from opencompass.utils.text_postprocessors import first_option_postprocess
# 指定只使用生命科学相关的子集
mmlu_life_science_subsets = [
'anatomy', # 解剖学
'clinical_knowledge', # 临床知识
'professional_medicine', # 专业医学
'medical_genetics', # 遗传学
'college_medicine', # 大学医学
'college_biology', # 大学生物学
]
mmlu_reader_cfg = dict(
input_columns=['input', 'A', 'B', 'C', 'D'],
output_column='target',
train_split='dev')
mmlu_datasets = []
for _name in mmlu_life_science_subsets:
_hint = f'There is a single choice question about {_name.replace("_", " ")}. Answer the question by replying A, B, C or D.'
mmlu_infer_cfg = dict(
ice_template=dict(
type=PromptTemplate,
template=dict(round=[
dict(
role='HUMAN',
prompt=f'{_hint}\nQuestion: {{input}}\nA. {{A}}\nB. {{B}}\nC. {{C}}\nD. {{D}}\nAnswer: '),
dict(role='BOT', prompt='{target}\n')
]),
),
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin='</E>',
round=[
dict(
role='HUMAN',
prompt=f'{_hint}\nQuestion: {{input}}\nA. {{A}}\nB. {{B}}\nC. {{C}}\nD. {{D}}\nAnswer: '
),
],
),
ice_token='</E>',
),
retriever=dict(type=FixKRetriever, fix_id_list=[0, 1, 2, 3, 4]),
inferencer=dict(type=GenInferencer),
)
mmlu_eval_cfg = dict(
evaluator=dict(type=AccwithDetailsEvaluator),
pred_postprocessor=dict(type=first_option_postprocess, options='ABCD'))
mmlu_datasets.append(
dict(
abbr=f'lukaemon_mmlu_{_name}',
type=MMLUDataset,
path='opencompass/mmlu',
name=_name,
reader_cfg=mmlu_reader_cfg,
infer_cfg=mmlu_infer_cfg,
eval_cfg=mmlu_eval_cfg,
))
del _name, _hint

View File

@ -0,0 +1,111 @@
from mmengine.config import read_base
from opencompass.openicl.icl_prompt_template import PromptTemplate
from opencompass.openicl.icl_retriever import ZeroRetriever
from opencompass.openicl.icl_inferencer import GenInferencer
from opencompass.openicl.icl_evaluator import AccEvaluator
from opencompass.datasets import MMLUDataset
from opencompass.utils.text_postprocessors import match_answer_pattern
from opencompass.evaluator import GenericLLMEvaluator
from opencompass.datasets import generic_llmjudge_postprocess
with read_base():
from .mmlu_lifescience_sets import mmlu_life_science_subsets
# None of the mmlu dataset in huggingface is correctly parsed, so we use our own dataset reader
# Please download the dataset from https://people.eecs.berkeley.edu/~hendrycks/data.tar
QUERY_TEMPLATE = """
Answer the following multiple choice question. The last line of your response should be of the following format: 'ANSWER: $LETTER' (without quotes) where LETTER is one of ABCD.
{input}
A) {A}
B) {B}
C) {C}
D) {D}
""".strip()
GRADER_TEMPLATE = """
Please as a grading expert, judge whether the final answers given by the candidates below are consistent with the standard answers, that is, whether the candidates answered correctly.
Here are some evaluation criteria:
1. Please refer to the given standard answer. You don't need to re-generate the answer to the question because the standard answer has been given. You only need to judge whether the candidate's answer is consistent with the standard answer according to the form of the question. Don't try to answer the original question. You can assume that the standard answer is definitely correct.
2. Because the candidate's answer may be different from the standard answer in the form of expression, before making a judgment, please understand the question and the standard answer first, and then judge whether the candidate's answer is correct, but be careful not to try to answer the original question.
3. Some answers may contain multiple items, such as multiple-choice questions, multiple-select questions, fill-in-the-blank questions, etc. As long as the answer is the same as the standard answer, it is enough. For multiple-select questions and multiple-blank fill-in-the-blank questions, the candidate needs to answer all the corresponding options or blanks correctly to be considered correct.
4. Some answers may be expressed in different ways, such as some answers may be a mathematical expression, some answers may be a textual description, as long as the meaning expressed is the same. And some formulas are expressed in different ways, but they are equivalent and correct.
Please judge whether the following answers are consistent with the standard answer based on the above criteria. Grade the predicted answer of this new question as one of:
A: CORRECT
B: INCORRECT
Just return the letters "A" or "B", with no text around it.
Here is your task. Simply reply with either CORRECT, INCORRECT. Don't apologize or correct yourself if there was a mistake; we are just trying to grade the answer.
<Original Question Begin>: {input}\n A) {A}\n B) {B}\n C) {C}\n D) {D}\n<Original Question End>\n\n
<Gold Target Begin>: \n{target}\n<Gold Target End>\n\n
<Predicted Answer Begin>: \n{prediction}\n<Predicted End>\n\n
Judging the correctness of candidates' answers:
""".strip()
mmlu_reader_cfg = dict(
input_columns=['input', 'A', 'B', 'C', 'D'],
output_column='target',
train_split='dev',
)
mmlu_datasets = []
for name in mmlu_life_science_subsets:
mmlu_infer_cfg = dict(
prompt_template=dict(
type=PromptTemplate,
template=dict(
round=[
dict(role='HUMAN', prompt=QUERY_TEMPLATE),
],
),
),
retriever=dict(type=ZeroRetriever),
inferencer=dict(type=GenInferencer),
)
mmlu_eval_cfg = dict(
evaluator=dict(
type=GenericLLMEvaluator,
prompt_template=dict(
type=PromptTemplate,
template=dict(
begin=[
dict(
role='SYSTEM',
fallback_role='HUMAN',
prompt="You are a helpful assistant who evaluates the correctness and quality of models' outputs.",
)
],
round=[
dict(role='HUMAN', prompt=GRADER_TEMPLATE),
],
),
),
dataset_cfg=dict(
type=MMLUDataset,
path='opencompass/mmlu',
name=name,
reader_cfg=mmlu_reader_cfg,
),
judge_cfg=dict(),
dict_postprocessor=dict(type=generic_llmjudge_postprocess),
),
pred_role='BOT',
)
mmlu_datasets.append(
dict(
abbr=f'lukaemon_mmlu_{name}',
type=MMLUDataset,
path='opencompass/mmlu',
name=name,
reader_cfg=mmlu_reader_cfg,
infer_cfg=mmlu_infer_cfg,
eval_cfg=mmlu_eval_cfg,
mode='singlescore',
)
)

View File

@ -0,0 +1,8 @@
mmlu_life_science_subsets = [
'anatomy', # 解剖学
'clinical_knowledge', # 临床知识
'professional_medicine', # 专业医学
'medical_genetics', # 遗传学
'college_medicine', # 大学医学
'college_biology', # 大学生物学
]

View File

@ -0,0 +1,62 @@
import re
from typing import List
from datasets import Dataset, DatasetDict, load_dataset
from opencompass.datasets.base import BaseDataset
from opencompass.registry import LOAD_DATASET
# 预编译的多选题正则,按 PEP-8 每行 < 79 字符
_PATTERN_MC = (
r'^(?P<stem>.*?)' # 题干
r'(?:A\.)\s*(?P<A>.*?)\s*' # 选项 A
r'B\.\s*(?P<B>.*?)\s*' # 选项 B
r'C\.\s*(?P<C>.*?)\s*' # 选项 C
r'D\.\s*(?P<D>.*?)' # 选项 D
r'Answer:' # 答案分隔符
)
@LOAD_DATASET.register_module()
class SciEvalDataset(BaseDataset):
"""Biology multiple-choice subset of SciEval."""
@staticmethod
def load(path: str, name: str, **kwargs) -> DatasetDict:
dataset = DatasetDict()
for split in ('test', ):
raw_iter = load_dataset(
path,
name=name,
split=split,
streaming=True,
)
examples: List[dict] = []
for ex in raw_iter:
if (ex.get('category') != 'biology'
or ex.get('type') != 'multiple-choice'):
continue
ans_list = ex.get('answer') or ex.get('answers') or []
if not ans_list:
continue
target = ans_list[0]
match = re.search(_PATTERN_MC, ex.get('question', ''), re.S)
if not match:
continue
examples.append({
'input': match.group('stem').strip(),
'A': match.group('A').strip(),
'B': match.group('B').strip(),
'C': match.group('C').strip(),
'D': match.group('D').strip(),
'target': target,
})
dataset[split] = Dataset.from_list(examples)
return dataset

View File

@ -99,6 +99,7 @@ from .MedXpertQA import * # noqa: F401, F403
from .mgsm import * # noqa: F401, F403
from .mmlu import * # noqa: F401, F403
from .mmlu_cf import * # noqa: F401, F403
from .mmlu_lifescience import * # noqa: F401, F403
from .mmlu_pro import * # noqa: F401, F403
from .MMLUArabic import * # noqa: F401, F403
from .mmmlu import * # noqa: F401, F403
@ -127,6 +128,7 @@ from .ruler import * # noqa: F401, F403
from .safety import * # noqa: F401, F403
from .scibench import ScibenchDataset, scibench_postprocess # noqa: F401, F403
from .scicode import * # noqa: F401, F403
from .SciEval_lifescience import SciEvalDataset # noqa: F401
from .simpleqa import * # noqa: F401, F403
from .siqa import * # noqa: F401, F403
from .smolinstruct import * # noqa: F401, F403

View File

@ -0,0 +1,153 @@
import csv
import json
import os.path as osp
from os import environ
from datasets import Dataset, DatasetDict
from opencompass.registry import LOAD_DATASET
from opencompass.utils import get_data_path
from .base import BaseDataset
@LOAD_DATASET.register_module()
class MMLUDataset(BaseDataset):
@staticmethod
def load(path: str, name: str, **kwargs):
path = get_data_path(path)
dataset = DatasetDict()
if environ.get('DATASET_SOURCE') == 'ModelScope':
from modelscope import MsDataset
for split in ['dev', 'test']:
# 从 ModelScope 加载数据
ms_dataset = MsDataset.load(path,
subset_name=name,
split=split)
dataset_list = []
for line in ms_dataset:
dataset_list.append({
'input': line['question'],
'A': line['choices'][0],
'B': line['choices'][1],
'C': line['choices'][2],
'D': line['choices'][3],
'target': 'ABCD'[line['answer']],
})
dataset[split] = Dataset.from_list(dataset_list)
else:
for split in ['dev', 'test']:
raw_data = []
filename = osp.join(path, split, f'{name}_{split}.csv')
with open(filename, encoding='utf-8') as f:
reader = csv.reader(f)
for row in reader:
assert len(row) == 6
raw_data.append({
'input': row[0],
'A': row[1],
'B': row[2],
'C': row[3],
'D': row[4],
'target': row[5],
})
dataset[split] = Dataset.from_list(raw_data)
return dataset
class MMLUDatasetClean(BaseDataset):
# load the contamination annotations of CEval from
# https://github.com/liyucheng09/Contamination_Detector
@staticmethod
def load_contamination_annotations(path, split='val'):
import requests
assert split == 'test', 'We only use test set for MMLU'
if environ.get('DATASET_SOURCE') == 'ModelScope':
from modelscope.utils.config_ds import MS_DATASETS_CACHE
annotation_cache_path = osp.join(
MS_DATASETS_CACHE,
f'MMLU_{split}_contamination_annotations.json')
link_of_annotations = 'https://modelscope.cn/datasets/opencompass/Contamination_Detector/resolve/master/mmlu_annotations.json' # noqa
else:
annotation_cache_path = osp.join(
path, split, f'MMLU_{split}_contamination_annotations.json')
link_of_annotations = 'https://github.com/liyucheng09/Contamination_Detector/releases/download/v0.1.1rc2/mmlu_annotations.json' # noqa
if osp.exists(annotation_cache_path):
with open(annotation_cache_path, 'r') as f:
annotations = json.load(f)
return annotations
annotations = json.loads(requests.get(link_of_annotations).text)
with open(annotation_cache_path, 'w') as f:
json.dump(annotations, f)
return annotations
@staticmethod
def load(path: str, name: str):
path = get_data_path(path)
dataset = DatasetDict()
if environ.get('DATASET_SOURCE') == 'ModelScope':
for split in ['dev', 'test']:
from modelscope import MsDataset
# 从 ModelScope 加载数据
ms_dataset = MsDataset.load(path,
subset_name=name,
split=split)
if split == 'test':
annotations = \
MMLUDatasetClean.load_contamination_annotations(
path, split)
dataset_list = []
for row_index, line in enumerate(ms_dataset):
item = {
'input': line['question'],
'A': line['choices'][0],
'B': line['choices'][1],
'C': line['choices'][2],
'D': line['choices'][3],
'target': 'ABCD'[line['answer']],
}
if split == 'test':
row_id = f'{name} {row_index}'
if row_id in annotations:
is_clean = annotations[row_id][0]
else:
is_clean = 'not labeled'
item['is_clean'] = is_clean
dataset_list.append(item)
dataset[split] = Dataset.from_list(dataset_list)
else:
for split in ['dev', 'test']:
raw_data = []
filename = osp.join(path, split, f'{name}_{split}.csv')
if split == 'test':
annotations = \
MMLUDatasetClean.load_contamination_annotations(
path, split)
with open(filename, encoding='utf-8') as f:
reader = csv.reader(f)
for row_index, row in enumerate(reader):
assert len(row) == 6
item = {
'input': row[0],
'A': row[1],
'B': row[2],
'C': row[3],
'D': row[4],
'target': row[5],
}
if split == 'test':
row_id = f'{name} {row_index}'
if row_id in annotations:
is_clean = annotations[row_id][0]
else:
is_clean = 'not labeled'
item['is_clean'] = is_clean
raw_data.append(item)
dataset[split] = Dataset.from_list(raw_data)
return dataset