Qwen1.5-0.5B-Chat

MERA Created at 22.09.2024 21:18
0.217
The overall result
570
Place in the rating
Weak tasks:
403
RWSD
565
PARus
583
RCB
590
MultiQ
578
ruWorldTree
563
ruOpenBookQA
567
CheGeKa
556
ruMMLU
497
ruHateSpeech
566
ruDetox
526
ruHHH
554
ruTiE
505
USE
539
MathLogicQA
486
ruMultiAr
489
SimpleAr
321
LCS
580
BPS
557
ruModAr
598
MaMuRAMu
+16
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.1 Accuracy
RCB 0.333 / 0.263 Accuracy F1 macro
USE 0.055 Grade norm
RWSD 0.504 Accuracy
PARus 0.49 Accuracy
ruTiE 0.5 Accuracy
MultiQ 0.088 / 0.026 F1 Exact match
CheGeKa 0.007 / 0 F1 Exact match
ruModAr 0.005 Exact match
MaMuRAMu 0.268 Accuracy
ruMultiAr 0.09 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.275 Accuracy
ruWorldTree 0.303 / 0.279 Accuracy F1 macro
ruOpenBookQA 0.345 / 0.304 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 0.476 Accuracy
ruMMLU 0.298 Accuracy
SimpleAr 0.442 Exact match
ruHumanEval 0 / 0 / 0 Pass@k
ruHHH 0.478
ruHateSpeech 0.521
ruDetox 0.053
ruEthics
Correct God Ethical
Virtue -0.012 -0.003 -0.044
Law -0.008 0.003 -0.047
Moral -0.007 0.002 -0.041
Justice 0.014 0.002 -0.034
Utilitarianism -0.003 0.011 -0.026

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:

Qwen1.5-0.5B-Chat

Model size

464.0M

Model type:

Opened

SFT

Additional links:

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

Architecture description:

Qwen1.5-0.5B-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"]; \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, ruworldtree, ruopenbookqa, rumultiar, use, rwsd, mamuramu, multiq, rumodar, ruethics, ruhumaneval, rucodeeval, rummlu, ruhhh - 32768 \nrutie - 2000

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
Qwen1.5-0.5B-Chat
MERA
0.033 0 0.2 0.1 0 0.067 0 - 0 0 0 0 0 0 0.033 0.217 0.033 0 0.033 0.033 0 0 0.033 0 0 0.1 0.133 0.033 0.033 0.133 0.133
Model, team Honest Helpful Harmless
Qwen1.5-0.5B-Chat
MERA
0.492 0.475 0.466
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
Qwen1.5-0.5B-Chat
MERA
0.281 0.223 0.336 0.325 0.346 0.303 0.301 0.273 0.306 0.238 0.237 0.27 0.32 0.306 0.287 0.301 0.27 0.229 0.333 0.313 0.263 0.251 0.36 0.329 0.232 0.34 0.275 0.333 0.237 0.36 0.32 0.413 0.301 0.364 0.336 0.226 0.3 0.332 0.291 0.31 0.364 0.342 0.317 0.286 0.316 0.343 0.319 0.27 0.252 0.29 0.38 0.359 0.331 0.294 0.32 0.339 0.342
Model, team SIM FL STA
Qwen1.5-0.5B-Chat
MERA
0.635 0.498 0.292
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
Qwen1.5-0.5B-Chat
MERA
0.267 0.248 0.217 0.278 0.303 0.19 0.362 0.228 0.288 0.231 0.346 0.217 0.217 0.248 0.316 0.136 0.383 0.333 0.246 0.316 0.281 0.237 0.133 0.249 0.311 0.258 0.295 0.193 0.404 0.2 0.4 0.154 0.125 0.281 0.167 0.357 0.2 0.311 0.228 0.215 0.302 0.302 0.311 0.267 0.328 0.267 0.276 0.25 0.338 0.263 0.378 0.275 0.456 0.208 0.256 0.228 0.322
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-0.5B-Chat
MERA
-0.012 -0.008 -0.007 0.014 -0.003
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-0.5B-Chat
MERA
-0.003 0.003 0.002 0.002 0.011
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-0.5B-Chat
MERA
-0.044 -0.047 -0.041 -0.034 -0.026
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-0.5B-Chat
MERA
0.546 0.543 0.529 0.649 0.429 0.393