Qwen1.5-4B-Chat

MERA Created at 22.09.2024 21:31
0.314
The overall result
462
Place in the rating
Weak tasks:
315
RWSD
494
PARus
499
RCB
482
ruEthics
400
MultiQ
474
ruWorldTree
448
ruOpenBookQA
506
CheGeKa
446
ruMMLU
486
ruHateSpeech
548
ruDetox
504
ruHHH
378
ruTiE
483
USE
348
MathLogicQA
343
ruMultiAr
333
SimpleAr
396
LCS
504
BPS
452
ruModAr
476
MaMuRAMu
+17
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.088 Accuracy
RCB 0.393 / 0.351 Accuracy F1 macro
USE 0.06 Grade norm
RWSD 0.523 Accuracy
PARus 0.528 Accuracy
ruTiE 0.563 Accuracy
MultiQ 0.266 / 0.18 F1 Exact match
CheGeKa 0.013 / 0 F1 Exact match
ruModAr 0.194 Exact match
MaMuRAMu 0.432 Accuracy
ruMultiAr 0.215 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.365 Accuracy
ruWorldTree 0.596 / 0.596 Accuracy F1 macro
ruOpenBookQA 0.548 / 0.546 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 0.549 Accuracy
ruMMLU 0.4 Accuracy
SimpleAr 0.92 Exact match
ruHumanEval 0 / 0 / 0 Pass@k
ruHHH 0.483
ruHateSpeech 0.525
ruDetox 0.055
ruEthics
Correct God Ethical
Virtue 0.094 0.226 0.09
Law 0.068 0.218 0.113
Moral 0.104 0.222 0.107
Justice 0.038 0.206 0.067
Utilitarianism 0.068 0.154 0.095

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
4.43.2
The number of GPUs and their type
1 x NVIDIA H100 80GB HBM3
Architecture
vllm

Team:

MERA

Name of the ML model:

Qwen1.5-4B-Chat

Model size

4.0B

Model type:

Opened

SFT

Additional links:

https://qwenlm.github.io/blog/qwen1.5/

Architecture description:

Qwen1.5-4B-chat is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, mixture of sliding window attention and full attention.

Description of the training:

We pretrained the model with a large amount of data with next-token prediction task, and we post-trained the model with both Direct Policy Optimization (DPO) and Proximal Policy Optimization (PPO).

Pretrain data:

Human conversations data formatted in ChatML-style and Qwen reward datasets, which are manually annotated Qwen models' answers.

License:

tongyi-qianwen-research

Inference parameters

Generation Parameters:
simplear - do_sample=false;until=["\n"]; \nchegeka - 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; \nrudetox - do_sample=false;until=["\n"];

The size of the context:
simplear, bps, lcs, chegeka, mathlogicqa, ruworldtree, ruopenbookqa, rumultiar, use, rwsd, mamuramu, multiq, rumodar, ruethics, ruhhh, ruhumaneval, rucodeeval, parus, rcb, rudetox, ruhatespeech, rummlu - 32768 \nrutie - 5000 \nrutie - 10000

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
Qwen1.5-4B-Chat
MERA
0.033 0.067 0.267 0.133 0 0.1 0 - 0 0 0 0 0 0 0.1 0.167 0.033 0.033 0 0.033 0 0 0.033 0 0.1 0.042 0.133 0.133 0.1 0.1 0.133
Model, team Honest Helpful Harmless
Qwen1.5-4B-Chat
MERA
0.443 0.525 0.483
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
Qwen1.5-4B-Chat
MERA
0.341 0.331 0.48 0.581 0.415 0.567 0.485 0.479 0.355 0.399 0.263 0.413 0.29 0.537 0.421 0.439 0.45 0.319 0.356 0.435 0.25 0.374 0.25 0.41 0.304 0.36 0.334 0.472 0.588 0.37 0.41 0.628 0.46 0.515 0.4 0.389 0.35 0.458 0.351 0.404 0.449 0.349 0.4 0.395 0.444 0.44 0.436 0.319 0.372 0.377 0.43 0.527 0.41 0.433 0.45 0.473 0.425
Model, team SIM FL STA
Qwen1.5-4B-Chat
MERA
0.845 0.502 0.156
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
Qwen1.5-4B-Chat
MERA
0.311 0.426 0.317 0.463 0.382 0.466 0.362 0.351 0.385 0.431 0.603 0.367 0.333 0.426 0.363 0.333 0.561 0.289 0.211 0.491 0.211 0.492 0.378 0.385 0.622 0.53 0.449 0.386 0.702 0.333 0.689 0.449 0.339 0.491 0.47 0.304 0.4 0.6 0.368 0.354 0.424 0.349 0.6 0.444 0.793 0.622 0.414 0.341 0.646 0.632 0.622 0.304 0.519 0.558 0.302 0.298 0.544
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-4B-Chat
MERA
0.094 0.068 0.104 0.038 0.068
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-4B-Chat
MERA
0.226 0.218 0.222 0.206 0.154
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-4B-Chat
MERA
0.09 0.113 0.107 0.067 0.095
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-4B-Chat
MERA
0.556 0.457 0.529 0.568 0.571 0.475