Qwen1.5-7B-Chat

MERA Created at 22.09.2024 21:36
0.369
The overall result
368
Place in the rating
Weak tasks:
130
RWSD
394
PARus
391
RCB
370
ruEthics
369
MultiQ
370
ruWorldTree
391
ruOpenBookQA
381
CheGeKa
385
ruMMLU
357
ruHateSpeech
432
ruDetox
303
ruHHH
324
ruTiE
392
ruHumanEval
400
USE
363
MathLogicQA
348
ruMultiAr
323
SimpleAr
561
LCS
384
BPS
380
ruModAr
397
MaMuRAMu
270
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.04 Accuracy
RCB 0.484 / 0.485 Accuracy F1 macro
USE 0.078 Grade norm
RWSD 0.585 Accuracy
PARus 0.648 Accuracy
ruTiE 0.61 Accuracy
MultiQ 0.287 / 0.147 F1 Exact match
CheGeKa 0.038 / 0.019 F1 Exact match
ruModAr 0.383 Exact match
MaMuRAMu 0.508 Accuracy
ruMultiAr 0.209 Exact match
ruCodeEval 0.002 / 0.008 / 0.012 Pass@k
MathLogicQA 0.359 Accuracy
ruWorldTree 0.745 / 0.748 Accuracy F1 macro
ruOpenBookQA 0.635 / 0.637 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 0.81 Accuracy
ruMMLU 0.445 Accuracy
SimpleAr 0.925 Exact match
ruHumanEval 0.001 / 0.003 / 0.006 Pass@k
ruHHH 0.629
ruHateSpeech 0.626
ruDetox 0.094
ruEthics
Correct God Ethical
Virtue 0.188 0.189 0.251
Law 0.188 0.186 0.244
Moral 0.216 0.184 0.261
Justice 0.127 0.153 0.204
Utilitarianism 0.143 0.167 0.203

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
6ec745e0
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
hf

Team:

MERA

Name of the ML model:

Qwen1.5-7B-Chat

Model size

7.7B

Model type:

Opened

SFT

Additional links:

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

Architecture description:

Qwen1.5-7B-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, 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
Qwen1.5-7B-Chat
MERA
0.133 0.167 0.367 0.1 0 0.133 0 - 0 0 0 0 0 0 0.133 0.233 0.033 0.033 0 0 0 0.033 0 0.033 0 0.158 0.133 0.033 0.067 0.033 0.133
Model, team Honest Helpful Harmless
Qwen1.5-7B-Chat
MERA
0.59 0.593 0.707
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
Qwen1.5-7B-Chat
MERA
0.393 0.386 0.487 0.645 0.493 0.622 0.553 0.518 0.444 0.448 0.281 0.365 0.29 0.481 0.498 0.486 0.52 0.41 0.256 0.565 0.237 0.45 0.3 0.422 0.295 0.43 0.351 0.519 0.616 0.4 0.52 0.62 0.417 0.646 0.483 0.406 0.34 0.545 0.331 0.443 0.566 0.441 0.407 0.398 0.564 0.412 0.588 0.322 0.358 0.425 0.44 0.654 0.426 0.479 0.58 0.642 0.554
Model, team SIM FL STA
Qwen1.5-7B-Chat
MERA
0.248 0.498 0.808
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
Qwen1.5-7B-Chat
MERA
0.378 0.485 0.433 0.491 0.395 0.603 0.448 0.544 0.404 0.508 0.564 0.508 0.325 0.527 0.415 0.333 0.617 0.467 0.404 0.579 0.316 0.576 0.511 0.503 0.578 0.515 0.513 0.421 0.719 0.556 0.711 0.551 0.482 0.579 0.424 0.429 0.444 0.756 0.439 0.462 0.555 0.444 0.6 0.489 0.776 0.689 0.586 0.545 0.631 0.632 0.622 0.333 0.633 0.584 0.465 0.409 0.578
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-7B-Chat
MERA
0.188 0.188 0.216 0.127 0.143
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-7B-Chat
MERA
0.189 0.186 0.184 0.153 0.167
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-7B-Chat
MERA
0.251 0.244 0.261 0.204 0.203
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-7B-Chat
MERA
0.704 0.6 0.353 0.676 0.286 0.59