Qwen1.5-32B

MERA Created at 22.09.2024 21:29
0.484
The overall result
172
Place in the rating
Weak tasks:
185
RWSD
179
PARus
214
RCB
340
ruEthics
222
MultiQ
190
ruWorldTree
214
ruOpenBookQA
226
CheGeKa
212
ruMMLU
142
ruHateSpeech
273
ruDetox
192
ruHHH
218
ruTiE
282
ruHumanEval
193
USE
211
MathLogicQA
115
ruMultiAr
155
SimpleAr
146
LCS
313
BPS
127
ruModAr
216
MaMuRAMu
278
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.14 Accuracy
RCB 0.541 / 0.445 Accuracy F1 macro
USE 0.184 Grade norm
RWSD 0.565 Accuracy
PARus 0.866 Accuracy
ruTiE 0.696 Accuracy
MultiQ 0.395 / 0.222 F1 Exact match
CheGeKa 0.143 / 0.106 F1 Exact match
ruModAr 0.635 Exact match
MaMuRAMu 0.703 Accuracy
ruMultiAr 0.341 Exact match
ruCodeEval 0.001 / 0.003 / 0.006 Pass@k
MathLogicQA 0.442 Accuracy
ruWorldTree 0.931 / 0.932 Accuracy F1 macro
ruOpenBookQA 0.825 / 0.825 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 0.903 Accuracy
ruMMLU 0.605 Accuracy
SimpleAr 0.982 Exact match
ruHumanEval 0.007 / 0.024 / 0.043 Pass@k
ruHHH 0.736
ruHateSpeech 0.789
ruDetox 0.172
ruEthics
Correct God Ethical
Virtue 0.171 0.201 0.211
Law 0.181 0.186 0.224
Moral 0.198 0.214 0.232
Justice 0.156 0.175 0.204
Utilitarianism 0.155 0.194 0.183

Information about the submission

Mera version
v.1.2.0
Torch Version
2.3.1
The version of the codebase
aec92f8
CUDA version
12.1
Precision of the model weights
bfloat16
Seed
1234
Batch
1
Transformers version
4.43.3
The number of GPUs and their type
4 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Qwen1.5-32B

Model size

32.5B

Model type:

Opened

Pretrain

Additional links:

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

Architecture description:

Qwen1.5-32B 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:

The model is trained using next-token prediction.

Pretrain data:

Dataset consists of around 3T tokens and includes public web documents, encyclopedia, books, codes, etc. Additionally, dataset is multilingual, with a significant portion of the data being in English and Chinese. 1. To increase the diversity of our data, the deduplication techniques were employed, including exact-match deduplication after normalization and fuzzy deduplication using MinHash and LSH algorithms. 2. To filter out low-quality data, a combination of rule-based and machine-learning-based methods were used. 3. Texts are manually sampled from various sources and reviewed to ensure their quality. 4. To further enhance the performance of the model, high-quality instruction data have been incorporated into the pretraining process.

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, rummlu, ruworldtree, ruopenbookqa, rumultiar, use, rwsd, mamuramu, multiq, rumodar, ruethics, ruhhh, rutie, ruhumaneval, rucodeeval - 32768

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
MERA
0.4 0.267 0.7 0.1 0.033 0.267 0 - 0.033 0 0 0 0.1 0 0.133 0.083 0.067 0.033 0 0.033 0.033 0.733 0.3 0.033 0 0.442 0.167 0.067 0.5 0.167 0.167
Model, team Honest Helpful Harmless
Qwen1.5-32B
MERA
0.639 0.729 0.845
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
MERA
0.474 0.542 0.763 0.803 0.722 0.821 0.767 0.688 0.688 0.673 0.421 0.492 0.39 0.667 0.75 0.676 0.71 0.667 0.389 0.695 0.226 0.737 0.45 0.63 0.42 0.67 0.458 0.574 0.722 0.43 0.7 0.818 0.613 0.879 0.634 0.62 0.44 0.771 0.47 0.557 0.798 0.625 0.538 0.568 0.757 0.486 0.804 0.393 0.436 0.585 0.46 0.814 0.7 0.676 0.83 0.8 0.772
Model, team SIM FL STA
Qwen1.5-32B
MERA
0.619 0.642 0.52
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
MERA
0.511 0.683 0.633 0.583 0.658 0.793 0.638 0.702 0.769 0.723 0.731 0.658 0.525 0.682 0.678 0.605 0.692 0.733 0.526 0.807 0.702 0.814 0.689 0.645 0.644 0.712 0.718 0.596 0.912 0.822 0.822 0.821 0.67 0.86 0.606 0.643 0.711 0.756 0.614 0.631 0.8 0.762 0.733 0.844 0.879 0.778 0.948 0.818 0.646 0.947 0.733 0.681 0.759 0.636 0.558 0.596 0.789
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B
MERA
0.171 0.181 0.198 0.156 0.155
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B
MERA
0.201 0.186 0.214 0.175 0.194
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-32B
MERA
0.211 0.224 0.232 0.204 0.183
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-32B
MERA
0.778 0.743 0.824 0.838 0.286 0.852