Qwen2.5-Math-72B

MERA Created at 04.11.2024 18:01
0.505
The overall result
147
Place in the rating
Weak tasks:
202
RWSD
135
PARus
357
RCB
262
ruEthics
491
MultiQ
195
ruWorldTree
162
ruOpenBookQA
271
CheGeKa
125
ruMMLU
165
ruHateSpeech
318
ruDetox
273
ruHHH
399
ruTiE
255
ruHumanEval
151
USE
56
MathLogicQA
51
ruMultiAr
54
SimpleAr
216
LCS
82
BPS
52
ruModAr
151
MaMuRAMu
277
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.12 Accuracy
RCB 0.498 / 0.431 Accuracy F1 macro
USE 0.216 Grade norm
RWSD 0.558 Accuracy
PARus 0.884 Accuracy
ruTiE 0.553 Accuracy
MultiQ 0.191 / 0.118 F1 Exact match
CheGeKa 0.115 / 0.072 F1 Exact match
ruModAr 0.832 Exact match
MaMuRAMu 0.746 Accuracy
ruMultiAr 0.441 Exact match
ruCodeEval 0.001 / 0.003 / 0.006 Pass@k
MathLogicQA 0.727 Accuracy
ruWorldTree 0.93 / 0.929 Accuracy F1 macro
ruOpenBookQA 0.86 / 0.86 Accuracy F1 macro

Evaluation on open tasks:

Go to the ratings by subcategory

The table will scroll to the left

Task name Result Metric
BPS 0.984 Accuracy
ruMMLU 0.672 Accuracy
SimpleAr 0.996 Exact match
ruHumanEval 0.012 / 0.052 / 0.091 Pass@k
ruHHH 0.652
ruHateSpeech 0.777
ruDetox 0.149
ruEthics
Correct God Ethical
Virtue 0.214 0.364 0.312
Law 0.233 0.362 0.319
Moral 0.259 0.388 0.322
Justice 0.216 0.294 0.27
Utilitarianism 0.217 0.318 0.27

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
fea61e4
CUDA version
12.1
Precision of the model weights
bfloat16bfloat16
Seed
1234
Batch
1
Transformers version
4.45.1
The number of GPUs and their type
8 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Qwen2.5-Math-72B

Model size

72.7B

Model type:

Opened

Pretrain

Additional links:

https://qwenlm.github.io/blog/qwen2.5-math/

Architecture description:

-

Description of the training:

-

Pretrain data:

First, the Qwen2-Math base models are trained on a high-quality mathematical pre-training dataset called the Qwen Math Corpus v1, which contains approximately 700 billion tokens. Second, we train a math-specific reward model Qwen2-Math-RM, derived from Qwen2-Math-72B, to create the Qwen2-Math-Instruct models. This reward model is used to construct Supervised Fine-Tuning (SFT) data through Rejection Sampling. Third, leveraging the Qwen2-Math-72B-Instruct model, we synthesize additional high-quality mathematical pre-training data, which serves as the foundation for Qwen Math Corpus v2. This updated corpus contains over 1 trillion tokens and is used to pre-train the Qwen2.5-Math models. Lastly, similar to the process used for the Qwen2-Math-Instruct models, we construct the Qwen2.5-Math-RM and Qwen2.5-Math-Instruct models.

License:

apache-2.0

Inference parameters

Generation Parameters:
simplear - do_sample=false;until=["\n"]; \nchegeka - do_sample=false;until=["\n"]; \nrudetox - do_sample=false;until=["\n"]; \nrumultiar - do_sample=false;until=["\n"]; \nuse - do_sample=false;until=["\n","."]; \nmultiq - do_sample=false;until=["\n"]; \nrumodar - do_sample=false;until=["\n"]; \nruhumaneval - do_sample=true;until=["\nclass","\ndef","\n#","\nif","\nprint"];temperature=0.6; \nrucodeeval - do_sample=true;until=["\nclass","\ndef","\n#","\nif","\nprint"];temperature=0.6;

The size of the context:
simplear, bps, lcs, chegeka, mathlogicqa, parus, rcb, rudetox, ruhatespeech, rummlu, ruworldtree, ruopenbookqa, rumultiar, use, rwsd, mamuramu, multiq, rumodar, ruethics, ruhhh, ruhumaneval, rucodeeval, rutie - 4096

System prompt:
Реши задачу по инструкции ниже. Не давай никаких объяснений и пояснений к своему ответу. Не пиши ничего лишнего. Пиши только то, что указано в инструкции. Если по инструкции нужно решить пример, то напиши только числовой ответ без хода решения и пояснений. Если по инструкции нужно вывести букву, цифру или слово, выведи только его. Если по инструкции нужно выбрать один из вариантов ответа и вывести букву или цифру, которая ему соответствует, то выведи только эту букву или цифру, не давай никаких пояснений, не добавляй знаки препинания, только 1 символ в ответе. Если по инструкции нужно дописать код функции на языке Python, пиши сразу код, соблюдая отступы так, будто ты продолжаешь функцию из инструкции, не давай пояснений, не пиши комментарии, используй только аргументы из сигнатуры функции в инструкции, не пробуй считывать данные через функцию input. Не извиняйся, не строй диалог. Выдавай только ответ и ничего больше.

Ratings by subcategory

Metric: Grade Norm
Model, team 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 8_0 8_1 8_2 8_3 8_4
Qwen2.5-Math-72B
MERA
0.433 0.367 0.767 0.167 0.067 0.3 0.033 - 0.033 0 0 0 0.2 0.1 0.133 0.167 0 0 0 0.033 0.067 0.5 0.167 0.067 0.067 0.533 0.133 0.1 0.5 0.333 0.3
Model, team Honest Helpful Harmless
Qwen2.5-Math-72B
MERA
0.557 0.695 0.707
Model, team Anatomy Virology Astronomy Marketing Nutrition Sociology Management Philosophy Prehistory Human aging Econometrics Formal logic Global facts Jurisprudence Miscellaneous Moral disputes Business ethics Biology (college) Physics (college) Human Sexuality Moral scenarios World religions Abstract algebra Medicine (college) Machine learning Medical genetics Professional law PR Security studies Chemistry (школьная) Computer security International law Logical fallacies Politics Clinical knowledge Conceptual_physics Math (college) Biology (high school) Physics (high school) Chemistry (high school) Geography (high school) Professional medicine Electrical engineering Elementary mathematics Psychology (high school) Statistics (high school) History (high school) Math (high school) Professional accounting Professional psychology Computer science (college) World history (high school) Macroeconomics Microeconomics Computer science (high school) European history Government and politics
Qwen2.5-Math-72B
MERA
0.511 0.5 0.855 0.812 0.716 0.756 0.738 0.704 0.639 0.677 0.623 0.571 0.53 0.713 0.765 0.72 0.7 0.819 0.733 0.672 0.255 0.749 0.65 0.653 0.67 0.75 0.454 0.639 0.731 0.63 0.77 0.752 0.699 0.798 0.668 0.833 0.8 0.871 0.735 0.744 0.798 0.621 0.759 0.844 0.854 0.782 0.799 0.644 0.543 0.669 0.71 0.755 0.803 0.929 0.88 0.752 0.772
Model, team SIM FL STA
Qwen2.5-Math-72B
MERA
0.769 0.615 0.365
Model, team Anatomy Virology Astronomy Marketing Nutrition Sociology Managment Philosophy Pre-History Gerontology Econometrics Formal logic Global facts Jurisprudence Miscellaneous Moral disputes Business ethics Bilology (college) Physics (college) Human sexuality Moral scenarios World religions Abstract algebra Medicine (college) Machine Learning Genetics Professional law PR Security Chemistry (college) Computer security International law Logical fallacies Politics Clinical knowledge Conceptual physics Math (college) Biology (high school) Physics (high school) Chemistry (high school) Geography (high school) Professional medicine Electrical Engineering Elementary mathematics Psychology (high school) Statistics (high school) History (high school) Math (high school) Professional Accounting Professional psychology Computer science (college) World history (high school) Macroeconomics Microeconomics Computer science (high school) Europe History Government and politics
Qwen2.5-Math-72B
MERA
0.556 0.743 0.767 0.694 0.671 0.69 0.69 0.737 0.654 0.754 0.872 0.792 0.525 0.736 0.626 0.679 0.72 0.756 0.789 0.825 0.719 0.831 0.844 0.746 0.867 0.742 0.679 0.614 0.895 0.822 0.822 0.808 0.804 0.842 0.515 0.857 0.933 0.8 0.772 0.8 0.792 0.794 0.8 0.978 0.862 0.867 0.879 0.886 0.708 0.825 0.822 0.638 0.823 0.766 0.605 0.556 0.789
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B
MERA
0.214 0.233 0.259 0.216 0.217
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B
MERA
0.364 0.362 0.388 0.294 0.318
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B
MERA
0.312 0.319 0.322 0.27 0.27
Model, team Women Men LGBT Nationalities Migrants Other
Qwen2.5-Math-72B
MERA
0.787 0.714 0.882 0.73 0.429 0.836