Qwen1.5-32B-Chat

MERA Created at 22.09.2024 21:27
0.491
The overall result
161
Place in the rating
Weak tasks:
80
RWSD
130
PARus
315
RCB
111
ruEthics
187
MultiQ
175
ruWorldTree
188
ruOpenBookQA
249
CheGeKa
189
ruMMLU
120
ruHateSpeech
328
ruDetox
99
ruHHH
170
ruTiE
300
ruHumanEval
147
USE
155
MathLogicQA
91
ruMultiAr
99
SimpleAr
92
LCS
347
BPS
271
ruModAr
190
MaMuRAMu
+18
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.16 Accuracy
RCB 0.511 / 0.459 Accuracy F1 macro
USE 0.219 Grade norm
RWSD 0.615 Accuracy
PARus 0.886 Accuracy
ruTiE 0.74 Accuracy
MultiQ 0.431 / 0.264 F1 Exact match
CheGeKa 0.125 / 0.077 F1 Exact match
ruModAr 0.472 Exact match
MaMuRAMu 0.714 Accuracy
ruMultiAr 0.364 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.474 Accuracy
ruWorldTree 0.937 / 0.937 Accuracy F1 macro
ruOpenBookQA 0.848 / 0.848 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 0.863 Accuracy
ruMMLU 0.627 Accuracy
SimpleAr 0.99 Exact match
ruHumanEval 0.006 / 0.006 / 0.006 Pass@k
ruHHH 0.837
ruHateSpeech 0.8
ruDetox 0.144
ruEthics
Correct God Ethical
Virtue 0.399 0.414 0.495
Law 0.375 0.38 0.467
Moral 0.443 0.438 0.534
Justice 0.32 0.335 0.444
Utilitarianism 0.327 0.345 0.419

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
1195ac01
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
4 x NVIDIA H100 80GB HBM3
Architecture
openai-completions

Team:

MERA

Name of the ML model:

Qwen1.5-32B-Chat

Model size

32.5B

Model type:

Opened

SFT

Additional links:

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

Architecture description:

Qwen1.5-32B-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, chegeka, rudetox, rumultiar, use, multiq, rumodar, ruhumaneval, rucodeeval - 2048 \nbps, lcs, mathlogicqa, parus, rcb, ruhatespeech, rummlu, ruworldtree, ruopenbookqa, rwsd, mamuramu, ruethics - 8000 \nruhhh, rutie - 1000 \nruhhh - 2000

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
Qwen1.5-32B-Chat
MERA
0.4 0.333 0.8 0.2 0 0.133 0 - 0 0 0 0 0.033 0 0.067 0.233 0 0 0 0.067 0 0.767 0.4 0.1 0.133 0.475 0.067 0.1 0.467 0.5 0.5
Model, team Honest Helpful Harmless
Qwen1.5-32B-Chat
MERA
0.787 0.864 0.862
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
Qwen1.5-32B-Chat
MERA
0.496 0.506 0.743 0.795 0.68 0.796 0.757 0.74 0.654 0.664 0.535 0.556 0.44 0.722 0.775 0.65 0.68 0.646 0.433 0.71 0.431 0.702 0.47 0.601 0.446 0.63 0.463 0.583 0.714 0.5 0.71 0.793 0.681 0.859 0.657 0.632 0.42 0.81 0.523 0.537 0.818 0.614 0.572 0.594 0.792 0.583 0.824 0.404 0.454 0.613 0.52 0.772 0.679 0.668 0.79 0.806 0.808
Model, team SIM FL STA
Qwen1.5-32B-Chat
MERA
0.329 0.606 0.76
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
Qwen1.5-32B-Chat
MERA
0.511 0.752 0.6 0.565 0.711 0.793 0.69 0.667 0.692 0.692 0.679 0.642 0.525 0.705 0.696 0.654 0.701 0.778 0.561 0.825 0.825 0.864 0.711 0.663 0.844 0.788 0.692 0.649 0.93 0.756 0.844 0.846 0.679 0.842 0.53 0.679 0.689 0.733 0.561 0.615 0.829 0.81 0.733 0.778 0.879 0.844 0.879 0.841 0.646 0.895 0.844 0.71 0.797 0.636 0.535 0.591 0.778
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B-Chat
MERA
0.399 0.375 0.443 0.32 0.327
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B-Chat
MERA
0.414 0.38 0.438 0.335 0.345
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B-Chat
MERA
0.495 0.467 0.534 0.444 0.419
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-32B-Chat
MERA
0.843 0.686 0.765 0.784 0.571 0.836