Qwen1.5-1.8B-Chat

MERA Created at 22.09.2024 21:21
0.262
The overall result
528
Place in the rating
Weak tasks:
240
RWSD
585
PARus
567
RCB
485
ruEthics
564
MultiQ
545
ruWorldTree
547
ruOpenBookQA
577
CheGeKa
539
ruMMLU
470
ruHateSpeech
623
ruDetox
546
ruHHH
576
ruTiE
343
ruHumanEval
546
USE
450
MathLogicQA
416
ruMultiAr
439
SimpleAr
437
LCS
575
BPS
414
ruModAr
541
MaMuRAMu
301
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.082 Accuracy
RCB 0.342 / 0.287 Accuracy F1 macro
USE 0.043 Grade norm
RWSD 0.55 Accuracy
PARus 0.478 Accuracy
ruTiE 0.492 Accuracy
MultiQ 0.123 / 0.047 F1 Exact match
CheGeKa 0.006 / 0 F1 Exact match
ruModAr 0.292 Exact match
MaMuRAMu 0.343 Accuracy
ruMultiAr 0.155 Exact match
ruCodeEval 0.001 / 0.003 / 0.006 Pass@k
MathLogicQA 0.318 Accuracy
ruWorldTree 0.396 / 0.394 Accuracy F1 macro
ruOpenBookQA 0.385 / 0.381 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 0.481 Accuracy
ruMMLU 0.319 Accuracy
SimpleAr 0.732 Exact match
ruHumanEval 0.004 / 0.006 / 0.006 Pass@k
ruHHH 0.472
ruHateSpeech 0.536
ruDetox 0.036
ruEthics
Correct God Ethical
Virtue 0.098 0.095 0.063
Law 0.112 0.09 0.065
Moral 0.096 0.082 0.066
Justice 0.096 0.069 0.034
Utilitarianism 0.087 0.055 0.045

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:

Qwen1.5-1.8B-Chat

Model size

1.8B

Model type:

Opened

SFT

Additional links:

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

Architecture description:

Qwen1.5-1.8B-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
Qwen1.5-1.8B-Chat
MERA
0.033 0.033 0.233 0.033 0.033 0.033 0 - 0.033 0 0.033 0 0 0 0.067 0.117 0.033 0 0.1 0.033 0 0 0 0.033 0.067 0 0.133 0.133 0.067 0 0.1
Model, team Honest Helpful Harmless
Qwen1.5-1.8B-Chat
MERA
0.508 0.475 0.431
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
Qwen1.5-1.8B-Chat
MERA
0.244 0.223 0.309 0.385 0.34 0.383 0.359 0.344 0.358 0.408 0.219 0.286 0.33 0.426 0.347 0.332 0.33 0.264 0.156 0.366 0.248 0.304 0.29 0.266 0.241 0.38 0.321 0.333 0.278 0.28 0.4 0.545 0.319 0.535 0.302 0.338 0.33 0.319 0.272 0.296 0.343 0.268 0.29 0.321 0.303 0.278 0.392 0.267 0.273 0.289 0.29 0.397 0.354 0.328 0.42 0.382 0.275
Model, team SIM FL STA
Qwen1.5-1.8B-Chat
MERA
0.635 0.339 0.266
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
Qwen1.5-1.8B-Chat
MERA
0.289 0.406 0.333 0.37 0.237 0.328 0.379 0.351 0.288 0.369 0.346 0.333 0.342 0.411 0.246 0.309 0.402 0.311 0.175 0.281 0.246 0.305 0.244 0.367 0.356 0.258 0.333 0.368 0.561 0.444 0.4 0.436 0.384 0.386 0.288 0.268 0.178 0.444 0.246 0.231 0.322 0.317 0.378 0.356 0.483 0.422 0.31 0.227 0.462 0.316 0.533 0.319 0.532 0.338 0.349 0.287 0.4
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B-Chat
MERA
0.098 0.112 0.096 0.096 0.087
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B-Chat
MERA
0.095 0.09 0.082 0.069 0.055
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B-Chat
MERA
0.063 0.065 0.066 0.034 0.045
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-1.8B-Chat
MERA
0.583 0.514 0.588 0.486 0.429 0.492