ruGPT3-large

RussianNLP Created at 20.09.2024 13:19
0.205
The overall result
587
Place in the rating
Weak tasks:
509
RWSD
589
PARus
609
RCB
547
MultiQ
614
ruWorldTree
614
ruOpenBookQA
545
CheGeKa
598
ruMMLU
453
ruHateSpeech
590
ruDetox
532
ruHHH
526
ruTiE
405
ruHumanEval
371
USE
540
MathLogicQA
582
ruMultiAr
591
SimpleAr
224
LCS
543
BPS
565
ruModAr
599
MaMuRAMu
+17
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.12 Accuracy
RCB 0.32 / 0.314 Accuracy F1 macro
USE 0.087 Grade norm
RWSD 0.469 Accuracy
PARus 0.474 Accuracy
ruTiE 0.504 Accuracy
MultiQ 0.136 / 0.04 F1 Exact match
CheGeKa 0.009 / 0 F1 Exact match
ruModAr 0.002 Exact match
MaMuRAMu 0.266 Accuracy
ruMultiAr 0.006 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.275 Accuracy
ruWorldTree 0.24 / 0.199 Accuracy F1 macro
ruOpenBookQA 0.248 / 0.229 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 0.511 Accuracy
ruMMLU 0.25 Accuracy
SimpleAr 0.008 Exact match
ruHumanEval 0.001 / 0.005 / 0.006 Pass@k
ruHHH 0.478
ruHateSpeech 0.543
ruDetox 0.047
ruEthics
Correct God Ethical
Virtue -0.074 -0.06 -0.054
Law -0.069 -0.052 -0.061
Moral -0.078 -0.042 -0.042
Justice -0.066 -0.073 -0.056
Utilitarianism -0.066 -0.058 -0.048

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
float16
Seed
1234
Batch
1
Transformers version
4.43.3
The number of GPUs and their type
5 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

RussianNLP

Name of the ML model:

ruGPT3-large

Model size

760.3M

Model type:

Opened

Pretrain

Architecture description:

ruGPT-3 is a Russian counterpart of GPT-3 (Brown et al., 2020). The model architecture is created using the description by Brown et al. and the GPT-2 code base (Radford et al., 2019) from the Transformers library. ruGPT-3 is pretrained on the language modeling objective. The BBPE tokenizer with the vocabulary size of 5 · 104 tokens was used.

Description of the training:

The model was pretrained with sequence length 1024 using transformers lib by the SberDevices team on 80B tokens for 3 epochs. After that, the model was finetuned 1 epoch with sequence length 2048. Total training time for ruGPT3-large has taken 16 days on the cluster of 128 V100-SXM3 GPUs. The initial learning rate of 1e−4 and the Adam optimizer with β1 = 0.9, β2 = 0.99, and ϵ = 1e−8 were used.

Pretrain data:

450GB of texts. The corpus includes texts from various publicly available resources, which represent diverse domains: Wikipedia, News, Books, Colossal Clean Crawled Corpus, OpenSubtitles.

License:

MIT

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 - 2048

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
ruGPT3-large
RussianNLP
0.1 0.033 0.167 0.167 0 0.033 0 - 0.067 0.033 0 0.033 0 0 0.033 0.4 0 0 0.133 0 0.067 0 0.067 0 0 0.142 0.067 0.2 0.2 0.067 0.133
Model, team Honest Helpful Harmless
ruGPT3-large
RussianNLP
0.492 0.475 0.466
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
ruGPT3-large
RussianNLP
0.274 0.295 0.184 0.265 0.239 0.239 0.223 0.193 0.238 0.247 0.202 0.294 0.32 0.259 0.253 0.28 0.29 0.243 0.233 0.267 0.27 0.216 0.21 0.301 0.25 0.28 0.242 0.185 0.2 0.23 0.27 0.289 0.166 0.323 0.279 0.338 0.33 0.229 0.225 0.232 0.227 0.272 0.221 0.233 0.228 0.324 0.289 0.248 0.27 0.244 0.2 0.241 0.233 0.252 0.25 0.236 0.249
Model, team SIM FL STA
ruGPT3-large
RussianNLP
0.685 0.598 0.185
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
ruGPT3-large
RussianNLP
0.156 0.267 0.3 0.25 0.105 0.328 0.224 0.298 0.212 0.231 0.231 0.375 0.242 0.24 0.187 0.531 0.234 0.311 0.281 0.333 0.211 0.356 0.178 0.249 0.356 0.197 0.192 0.368 0.281 0.289 0.222 0.359 0.429 0.263 0.258 0.321 0.222 0.333 0.316 0.369 0.249 0.27 0.222 0.244 0.224 0.333 0.224 0.136 0.246 0.263 0.222 0.29 0.253 0.221 0.163 0.257 0.244
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-large
RussianNLP
-0.074 -0.069 -0.078 -0.066 -0.066
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-large
RussianNLP
-0.06 -0.052 -0.042 -0.073 -0.058
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-large
RussianNLP
-0.054 -0.061 -0.042 -0.056 -0.048
Model, team Women Men LGBT Nationalities Migrants Other
ruGPT3-large
RussianNLP
0.519 0.686 0.588 0.595 0.286 0.492