Qwen1.5-1.8B

MERA Created at 22.09.2024 21:24
0.267
The overall result
525
Place in the rating
Weak tasks:
444
RWSD
593
PARus
552
RCB
514
ruEthics
535
MultiQ
554
ruWorldTree
534
ruOpenBookQA
562
CheGeKa
536
ruMMLU
424
ruHateSpeech
618
ruDetox
516
ruHHH
519
ruTiE
211
ruHumanEval
561
USE
471
MathLogicQA
378
ruMultiAr
430
SimpleAr
412
LCS
561
BPS
403
ruModAr
538
MaMuRAMu
213
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.086 Accuracy
RCB 0.347 / 0.256 Accuracy F1 macro
USE 0.038 Grade norm
RWSD 0.492 Accuracy
PARus 0.47 Accuracy
ruTiE 0.504 Accuracy
MultiQ 0.152 / 0.063 F1 Exact match
CheGeKa 0.007 / 0 F1 Exact match
ruModAr 0.322 Exact match
MaMuRAMu 0.348 Accuracy
ruMultiAr 0.187 Exact match
ruCodeEval 0.023 / 0.073 / 0.098 Pass@k
MathLogicQA 0.309 Accuracy
ruWorldTree 0.39 / 0.374 Accuracy F1 macro
ruOpenBookQA 0.403 / 0.386 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 0.496 Accuracy
ruMMLU 0.32 Accuracy
SimpleAr 0.784 Exact match
ruHumanEval 0.048 / 0.131 / 0.177 Pass@k
ruHHH 0.478
ruHateSpeech 0.558
ruDetox 0.039
ruEthics
Correct God Ethical
Virtue 0.061 0.042 0.05
Law 0.071 0.021 0.013
Moral 0.069 0.039 0.039
Justice 0.035 0.007 0.028
Utilitarianism 0.034 0.012 0.048

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

Model size

1.8B

Model type:

Opened

Pretrain

Additional links:

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

Architecture description:

Qwen1.5-1.8B 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
Qwen1.5-1.8B
MERA
0 0.033 0.333 0.033 0 0 0 - 0 0 0 0 0 0 0.133 0.083 0.033 0 0.033 0.033 0 0 0 0 0 0.008 0.167 0.033 0.067 0.067 0.133
Model, team Honest Helpful Harmless
Qwen1.5-1.8B
MERA
0.492 0.492 0.448
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
MERA
0.267 0.337 0.276 0.415 0.379 0.373 0.32 0.331 0.324 0.359 0.272 0.294 0.35 0.463 0.355 0.335 0.39 0.257 0.189 0.336 0.253 0.292 0.28 0.335 0.214 0.27 0.299 0.417 0.294 0.22 0.47 0.62 0.319 0.485 0.309 0.363 0.36 0.323 0.212 0.315 0.318 0.235 0.366 0.297 0.277 0.25 0.348 0.3 0.291 0.323 0.34 0.371 0.356 0.353 0.39 0.327 0.285
Model, team SIM FL STA
Qwen1.5-1.8B
MERA
0.804 0.525 0.121
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
MERA
0.2 0.317 0.333 0.361 0.263 0.293 0.276 0.281 0.346 0.385 0.321 0.325 0.375 0.496 0.257 0.309 0.383 0.333 0.298 0.351 0.175 0.271 0.244 0.408 0.356 0.227 0.333 0.368 0.614 0.356 0.556 0.449 0.42 0.404 0.318 0.214 0.289 0.378 0.298 0.292 0.339 0.333 0.444 0.311 0.534 0.444 0.345 0.295 0.292 0.351 0.467 0.319 0.481 0.312 0.209 0.292 0.456
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B
MERA
0.061 0.071 0.069 0.035 0.034
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B
MERA
0.042 0.021 0.039 0.007 0.012
Model, team Virtue Law Moral Justice Utilitarianism
Qwen1.5-1.8B
MERA
0.05 0.013 0.039 0.028 0.048
Model, team Women Men LGBT Nationalities Migrants Other
Qwen1.5-1.8B
MERA
0.611 0.571 0.588 0.595 0.286 0.459