Qwen2.5-Math-72B-Instruct

MERA Created at 04.11.2024 18:01
0.507
The overall result
144
Place in the rating
Weak tasks:
95
RWSD
231
PARus
238
RCB
329
ruEthics
575
MultiQ
157
ruWorldTree
177
ruOpenBookQA
289
CheGeKa
130
ruMMLU
253
ruHateSpeech
284
ruDetox
191
ruHHH
280
ruTiE
147
ruHumanEval
143
USE
83
MathLogicQA
66
ruMultiAr
84
SimpleAr
233
LCS
302
BPS
86
ruModAr
161
MaMuRAMu
193
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.116 Accuracy
RCB 0.534 / 0.476 Accuracy F1 macro
USE 0.23 Grade norm
RWSD 0.604 Accuracy
PARus 0.826 Accuracy
ruTiE 0.653 Accuracy
MultiQ 0.114 / 0.061 F1 Exact match
CheGeKa 0.099 / 0.065 F1 Exact match
ruModAr 0.713 Exact match
MaMuRAMu 0.735 Accuracy
ruMultiAr 0.419 Exact match
ruCodeEval 0.046 / 0.194 / 0.329 Pass@k
MathLogicQA 0.648 Accuracy
ruWorldTree 0.943 / 0.943 Accuracy F1 macro
ruOpenBookQA 0.85 / 0.849 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 0.909 Accuracy
ruMMLU 0.667 Accuracy
SimpleAr 0.992 Exact match
ruHumanEval 0.126 / 0.376 / 0.518 Pass@k
ruHHH 0.736
ruHateSpeech 0.717
ruDetox 0.167
ruEthics
Correct God Ethical
Virtue 0.193 0.463 0.411
Law 0.24 0.481 0.402
Moral 0.222 0.453 0.44
Justice 0.207 0.441 0.41
Utilitarianism 0.163 0.367 0.352

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-Instruct

Model size

72.7B

Model type:

Opened

SFT

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-Instruct
MERA
0.433 0.267 0.8 0.133 0.067 0.2 0.033 - 0 0.033 0.1 0.033 0.067 0 0.1 0.35 0 0.033 0.033 0 0.067 0.6 0.267 0.033 0.1 0.525 0.167 0.1 0.6 0.433 0.333
Model, team Honest Helpful Harmless
Qwen2.5-Math-72B-Instruct
MERA
0.639 0.797 0.776
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-Instruct
MERA
0.526 0.464 0.842 0.842 0.703 0.751 0.757 0.704 0.657 0.659 0.675 0.595 0.44 0.722 0.768 0.639 0.69 0.764 0.667 0.664 0.394 0.719 0.66 0.659 0.616 0.71 0.435 0.639 0.718 0.62 0.73 0.752 0.687 0.778 0.653 0.833 0.75 0.874 0.695 0.734 0.823 0.61 0.69 0.785 0.828 0.778 0.794 0.604 0.525 0.654 0.7 0.768 0.792 0.903 0.89 0.752 0.767
Model, team SIM FL STA
Qwen2.5-Math-72B-Instruct
MERA
0.601 0.627 0.524
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-Instruct
MERA
0.511 0.762 0.75 0.657 0.724 0.707 0.655 0.719 0.692 0.692 0.885 0.767 0.533 0.674 0.626 0.568 0.71 0.8 0.789 0.825 0.614 0.746 0.911 0.74 0.867 0.758 0.641 0.596 0.895 0.778 0.844 0.795 0.696 0.807 0.53 0.893 0.889 0.8 0.754 0.8 0.82 0.746 0.778 0.978 0.879 0.867 0.862 0.909 0.692 0.86 0.889 0.565 0.823 0.74 0.651 0.567 0.789
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B-Instruct
MERA
0.193 0.24 0.222 0.207 0.163
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B-Instruct
MERA
0.463 0.481 0.453 0.441 0.367
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2.5-Math-72B-Instruct
MERA
0.411 0.402 0.44 0.41 0.352
Model, team Women Men LGBT Nationalities Migrants Other
Qwen2.5-Math-72B-Instruct
MERA
0.731 0.629 0.765 0.676 0.571 0.77