Megra 1.1 709M (SFT-4096)

Promezis Lab Created at 16.09.2026 20:22
0.22
The overall result
568
Place in the rating
Weak tasks:
448
RWSD
539
PARus
601
RCB
587
MultiQ
597
ruWorldTree
601
ruOpenBookQA
585
CheGeKa
594
ruMMLU
487
ruHateSpeech
528
ruDetox
572
ruHHH
559
ruTiE
554
USE
554
MathLogicQA
492
ruMultiAr
522
SimpleAr
40
LCS
367
BPS
499
ruModAr
605
MaMuRAMu
+16
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.244 Accuracy
RCB 0.329 / 0.207 Accuracy F1 macro
USE 0.04 Grade norm
RWSD 0.492 Accuracy
PARus 0.5 Accuracy
ruTiE 0.498 Accuracy
MultiQ 0.089 / 0.032 F1 Exact match
CheGeKa 0.006 / 0 F1 Exact match
ruModAr 0.096 Exact match
MaMuRAMu 0.258 Accuracy
ruMultiAr 0.079 Exact match
ruCodeEval 0 / 0 / 0 Pass@k
MathLogicQA 0.265 Accuracy
ruWorldTree 0.263 / 0.231 Accuracy F1 macro
ruOpenBookQA 0.26 / 0.241 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 0.844 Accuracy
ruMMLU 0.256 Accuracy
SimpleAr 0.348 Exact match
ruHumanEval 0 / 0 / 0 Pass@k
ruHHH 0.449
ruHateSpeech 0.525
ruDetox 0.065
ruEthics
Correct God Ethical
Virtue 0.037 -0.009 0.001
Law 0.01 -0.002 -0.02
Moral 0.017 -0.005 -0.025
Justice 0.009 -0.028 -0.023
Utilitarianism 0.024 0.044 0.012

Information about the submission

Mera version
v1.2.0
Torch Version
2.13.0
The version of the codebase
lm-eval 0.4.9.1
CUDA version
13.0
Precision of the model weights
FP32
Seed
1234
Batch
1
Transformers version
4.57.1
The number of GPUs and their type
1 x NVIDIA GeForce RTX 5080
Architecture
PromezisCorev4

Team:

Promezis Lab

Name of the ML model:

Megra 1.1 709M (SFT-4096)

Link to the ML model:

https://promezis.com/megra.html

Model size

709.0M

Model type:

Closed

SFT

Architecture description:

Megra - языковая модель Promezis Lab на архитектуре promeziscore v4. Плотная авторегрессионная Transformer-модель: 709 261 440 параметров, 46 слоёв, размерность скрытого состояния 1152. Используется Promezis Tokenizer со словарём 50 368 токенов. Контекст в представленном замере - 4096 токенов. Веса закрыты.

Description of the training:

Модель прошла предварительное обучение с нуля, продолженное предобучение и полно-параметрическую настройку SFT. Последующие этапы SFT включали арифметику, работу с контекстом, логический вывод и алгоритмические задачи с повторением ранее освоенных данных. Финальные веса получены равномерной интерполяцией двух последовательных SFT-чекпоинтов: 75% родительского и 25% дополнительного алгоритмического дообучения. Во всех задачах использовался один фиксированный набор весов.

Pretrain data:

Предобучение: текстовый корпус с русскоязычными материалами, математикой, кодом и предметными знаниями. SFT: инструкционные данные и синтетические проверяемые задания по арифметике, работе с контекстом, логическому выводу и алгоритмам. Использовались примеры SNLI/MNLI и повторение данных предыдущих этапов. Финальные алгоритмические данные проверялись на точные совпадения с входами закреплённой версии MERA.

License:

Проприетарная. Веса модели не предоставляются для публичного скачивания.

Inference parameters

Generation Parameters:
multiq - do_sample=false;until=[" \n"]; \nrucodeeval - do_sample=true;temperature=0.6;until=[" \nclass"," \ndef"," \n#"," \nif"," \nprint"]; \nruhumaneval - do_sample=true;temperature=0.6;until=[" \nclass"," \ndef"," \n#"," \nif"," \nprint"]; \nrumodar - 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","."]; \nsimplear - do_sample=false;until=[" \n"];

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
Megra 1.1 709M (SFT-4096)
Promezis Lab
0 0 0.233 0 0 0 0 - 0.033 0 0 0 0 0 0.033 0.083 0.067 0 0.067 0 0 0 0 0 0 0.042 0.067 0.133 0.133 0.067 0.2
Model, team Honest Helpful Harmless
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.393 0.508 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
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.222 0.241 0.303 0.278 0.219 0.269 0.291 0.277 0.259 0.287 0.193 0.238 0.32 0.231 0.268 0.295 0.27 0.208 0.2 0.214 0.272 0.216 0.21 0.231 0.196 0.26 0.25 0.194 0.229 0.2 0.27 0.273 0.276 0.242 0.204 0.308 0.27 0.265 0.238 0.251 0.263 0.254 0.262 0.263 0.238 0.236 0.275 0.248 0.316 0.23 0.25 0.236 0.297 0.277 0.26 0.267 0.264
Model, team SIM FL STA
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.452 0.626 0.379
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
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.178 0.257 0.267 0.352 0.289 0.293 0.293 0.228 0.231 0.262 0.231 0.258 0.225 0.24 0.246 0.185 0.308 0.289 0.298 0.211 0.228 0.271 0.267 0.314 0.156 0.242 0.231 0.211 0.404 0.311 0.378 0.231 0.313 0.333 0.242 0.339 0.333 0.2 0.368 0.154 0.208 0.317 0.2 0.222 0.207 0.289 0.259 0.295 0.231 0.298 0.311 0.203 0.253 0.221 0.326 0.164 0.311
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.037 0.01 0.017 0.009 0.024
Model, team Virtue Law Moral Justice Utilitarianism
Megra 1.1 709M (SFT-4096)
Promezis Lab
-0.009 -0.002 -0.005 -0.028 0.044
Model, team Virtue Law Moral Justice Utilitarianism
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.001 -0.02 -0.025 -0.023 0.012
Model, team Women Men LGBT Nationalities Migrants Other
Megra 1.1 709M (SFT-4096)
Promezis Lab
0.519 0.686 0.529 0.486 0.143 0.508