ruGPT3-medium

RussianNLP Created at 20.09.2024 13:03
0.207
The overall result
584
Place in the rating
Weak tasks:
523
RWSD
524
PARus
612
RCB
504
ruEthics
566
MultiQ
611
ruWorldTree
597
ruOpenBookQA
481
CheGeKa
601
ruMMLU
418
ruHateSpeech
594
ruDetox
531
ruHHH
525
ruTiE
404
ruHumanEval
436
USE
548
MathLogicQA
570
ruMultiAr
589
SimpleAr
286
LCS
568
BPS
570
ruModAr
597
MaMuRAMu
+18
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.108 Accuracy
RCB 0.317 / 0.314 Accuracy F1 macro
USE 0.07 Grade norm
RWSD 0.462 Accuracy
PARus 0.504 Accuracy
ruTiE 0.504 Accuracy
MultiQ 0.122 / 0.046 F1 Exact match
CheGeKa 0.016 / 0 F1 Exact match
ruModAr 0.001 Exact match
MaMuRAMu 0.27 Accuracy
ruMultiAr 0.009 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.27 Accuracy
ruWorldTree 0.246 / 0.233 Accuracy F1 macro
ruOpenBookQA 0.268 / 0.263 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 0.491 Accuracy
ruMMLU 0.246 Accuracy
SimpleAr 0.009 Exact match
ruHumanEval 0.001 / 0.005 / 0.006 Pass@k
ruHHH 0.478
ruHateSpeech 0.562
ruDetox 0.046
ruEthics
Correct God Ethical
Virtue 0.028 0.035 0.079
Law 0.014 0.021 0.093
Moral 0.018 0.025 0.072
Justice 0.04 0.047 0.1
Utilitarianism 0.055 0.064 0.076

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-medium

Model size

355.9M

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-medium has taken 16 days on the cluster of 64 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-medium
RussianNLP
0.1 0.033 0.233 0.167 0 0.067 0 - 0 0.067 0 0.033 0 0 0.033 0.367 0 0 0.067 0 0.067 0 0.033 0 0 0.083 0.1 0.133 0.067 0.033 0.067
Model, team Honest Helpful Harmless
ruGPT3-medium
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-medium
RussianNLP
0.259 0.247 0.191 0.265 0.235 0.279 0.233 0.219 0.259 0.283 0.211 0.27 0.25 0.259 0.232 0.269 0.27 0.257 0.222 0.229 0.277 0.246 0.28 0.318 0.25 0.29 0.229 0.185 0.2 0.21 0.23 0.281 0.252 0.283 0.242 0.248 0.27 0.265 0.192 0.212 0.212 0.305 0.255 0.202 0.217 0.347 0.289 0.219 0.252 0.244 0.26 0.232 0.277 0.223 0.21 0.212 0.275
Model, team SIM FL STA
ruGPT3-medium
RussianNLP
0.774 0.574 0.157
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-medium
RussianNLP
0.333 0.257 0.35 0.287 0.171 0.328 0.31 0.351 0.288 0.2 0.231 0.308 0.2 0.341 0.193 0.519 0.224 0.289 0.316 0.316 0.211 0.254 0.156 0.237 0.311 0.288 0.269 0.193 0.316 0.244 0.222 0.346 0.402 0.263 0.288 0.321 0.2 0.244 0.228 0.354 0.253 0.286 0.222 0.289 0.207 0.333 0.207 0.205 0.215 0.333 0.222 0.232 0.304 0.156 0.209 0.31 0.233
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-medium
RussianNLP
0.028 0.014 0.018 0.04 0.055
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-medium
RussianNLP
0.035 0.021 0.025 0.047 0.064
Model, team Virtue Law Moral Justice Utilitarianism
ruGPT3-medium
RussianNLP
0.079 0.093 0.072 0.1 0.076
Model, team Women Men LGBT Nationalities Migrants Other
ruGPT3-medium
RussianNLP
0.519 0.657 0.706 0.595 0.286 0.557