Qwen2-57B-A14B-Instruct

MERA Created at 22.09.2024 21:55
0.516
The overall result
129
Place in the rating
Weak tasks:
241
RWSD
88
PARus
237
RCB
21
ruEthics
128
MultiQ
156
ruWorldTree
120
ruOpenBookQA
162
CheGeKa
163
ruMMLU
77
ruHateSpeech
102
ruDetox
206
ruHHH
173
ruTiE
175
ruHumanEval
135
USE
123
MathLogicQA
146
ruMultiAr
111
SimpleAr
280
LCS
226
BPS
183
ruModAr
140
MaMuRAMu
157
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.108 Accuracy
RCB 0.534 / 0.456 Accuracy F1 macro
USE 0.241 Grade norm
RWSD 0.546 Accuracy
PARus 0.906 Accuracy
ruTiE 0.735 Accuracy
MultiQ 0.48 / 0.348 F1 Exact match
CheGeKa 0.19 / 0.144 F1 Exact match
ruModAr 0.513 Exact match
MaMuRAMu 0.752 Accuracy
ruMultiAr 0.322 Exact match
ruCodeEval 0.088 / 0.23 / 0.299 Pass@k
MathLogicQA 0.507 Accuracy
ruWorldTree 0.943 / 0.943 Accuracy F1 macro
ruOpenBookQA 0.885 / 0.885 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 0.947 Accuracy
ruMMLU 0.648 Accuracy
SimpleAr 0.989 Exact match
ruHumanEval 0.087 / 0.221 / 0.293 Pass@k
ruHHH 0.719
ruHateSpeech 0.826
ruDetox 0.29
ruEthics
Correct God Ethical
Virtue 0.457 0.444 0.458
Law 0.496 0.434 0.458
Moral 0.508 0.458 0.473
Justice 0.421 0.401 0.411
Utilitarianism 0.413 0.414 0.402

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
9b26db97
CUDA version
12.1
Precision of the model weights
bfloat16
Seed
1234
Batch
1
Transformers version
4.44.2
The number of GPUs and their type
8 x NVIDIA H100 80GB HBM3
Architecture
vllm

Team:

MERA

Name of the ML model:

Qwen2-57B-A14B-Instruct

Model size

57.4B

Model type:

Opened

SFT

Additional links:

https://arxiv.org/pdf/2407.10671

Architecture description:

Qwen2-57B-A14B-Instruct is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, RoPE and RMSNorm. As a substitute for the original FFN, the MoE FFN consists of n individual FFNs, each serving as an expert. Each token is directed to a specific expert $E_i$ for computation based on probabilities assigned by a gated network G.

Description of the training:

SFT and RLHF on the mix of manually and synthetically annotated samples after pertaining with next-token prediction task.

Pretrain data:

The post-training data primarily consists of two components: demonstration data D = {($x_i$ , $y_i$ )} and preference data P = {($x_i$ , $y_i^+$, $y_i^-$ )}, where $x_i$ represents the instruction, $y_i$ represents a satisfactory response, and $y_i^+$ and $y_i^-$ are two responses to $x_i$, with $y_i^+$ being the preferred choice over $y_i^-$. The set D is utilized in SFT, whereas P is employed in RLHF. We have assembled an extensive instruction dataset featuring more than 500,000 examples that cover skills such as instruction following, coding, mathematics, logical reasoning, role-playing, multilingualism, and safety.

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 - 32768 \nrutie - 5000

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.

Description of the template:
{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system \nYou are a helpful assistant.<|im_end|> \n' }}{% endif %}{{'<|im_start|>' + message['role'] + ' \n' + message['content'] + '<|im_end|>' + ' \n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant \n' }}{% endif %}

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
Qwen2-57B-A14B-Instruct
MERA
0.567 0.467 0.767 0.167 0.1 0.433 0.1 - 0.033 0.1 0.1 0.067 0.2 0 0.1 0.4 0.033 0.033 0 0 0.033 0.7 0.267 0.067 0.1 0.417 0.133 0.1 0.367 0.367 0.333
Model, team Honest Helpful Harmless
Qwen2-57B-A14B-Instruct
MERA
0.721 0.644 0.793
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
Qwen2-57B-A14B-Instruct
MERA
0.607 0.512 0.789 0.85 0.696 0.826 0.777 0.701 0.735 0.691 0.57 0.492 0.47 0.741 0.808 0.673 0.71 0.813 0.5 0.756 0.255 0.813 0.37 0.676 0.545 0.74 0.469 0.667 0.78 0.48 0.71 0.818 0.706 0.859 0.717 0.739 0.52 0.835 0.55 0.616 0.803 0.695 0.6 0.615 0.833 0.56 0.809 0.452 0.475 0.633 0.6 0.827 0.741 0.761 0.79 0.788 0.845
Model, team SIM FL STA
Qwen2-57B-A14B-Instruct
MERA
0.621 0.722 0.68
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
Qwen2-57B-A14B-Instruct
MERA
0.467 0.842 0.683 0.63 0.803 0.759 0.724 0.719 0.731 0.692 0.731 0.708 0.542 0.798 0.696 0.63 0.748 0.778 0.684 0.825 0.596 0.864 0.844 0.817 0.756 0.788 0.756 0.719 0.877 0.778 0.822 0.846 0.741 0.912 0.667 0.768 0.622 0.778 0.544 0.677 0.833 0.873 0.844 0.822 0.897 0.889 0.879 0.818 0.815 0.912 0.756 0.739 0.797 0.753 0.558 0.637 0.822
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-57B-A14B-Instruct
MERA
0.457 0.496 0.508 0.421 0.413
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-57B-A14B-Instruct
MERA
0.444 0.434 0.458 0.401 0.414
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-57B-A14B-Instruct
MERA
0.458 0.458 0.473 0.411 0.402
Model, team Women Men LGBT Nationalities Migrants Other
Qwen2-57B-A14B-Instruct
MERA
0.833 0.771 0.824 0.865 0.571 0.852