Meta-Llama-3-8B

MERA Created at 20.09.2024 16:01
0.445
The overall result
249
Place in the rating
Weak tasks:
187
RWSD
349
PARus
318
RCB
415
ruEthics
345
MultiQ
312
ruWorldTree
317
ruOpenBookQA
177
CheGeKa
313
ruMMLU
327
ruHateSpeech
471
ruDetox
298
ruHHH
359
ruTiE
93
ruHumanEval
333
USE
403
MathLogicQA
297
ruMultiAr
290
SimpleAr
259
LCS
399
BPS
99
ruModAr
287
MaMuRAMu
110
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.112 Accuracy
RCB 0.511 / 0.433 Accuracy F1 macro
USE 0.109 Grade norm
RWSD 0.565 Accuracy
PARus 0.69 Accuracy
ruTiE 0.582 Accuracy
MultiQ 0.308 / 0.141 F1 Exact match
CheGeKa 0.179 / 0.135 F1 Exact match
ruModAr 0.697 Exact match
MaMuRAMu 0.625 Accuracy
ruMultiAr 0.248 Exact match
ruCodeEval 0.19 / 0.405 / 0.488 Pass@k
MathLogicQA 0.339 Accuracy
ruWorldTree 0.804 / 0.803 Accuracy F1 macro
ruOpenBookQA 0.693 / 0.692 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 0.779 Accuracy
ruMMLU 0.51 Accuracy
SimpleAr 0.939 Exact match
ruHumanEval 0.225 / 0.404 / 0.482 Pass@k
ruHHH 0.629
ruHateSpeech 0.645
ruDetox 0.08
ruEthics
Correct God Ethical
Virtue 0.209 0.17 0.107
Law 0.199 0.146 0.117
Moral 0.198 0.171 0.109
Justice 0.177 0.176 0.096
Utilitarianism 0.166 0.128 0.095

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
4.42.3
The number of GPUs and their type
1 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Meta-Llama-3-8B

Model size

8.0B

Model type:

Opened

Pretrain

Additional links:

https://ai.meta.com/research/publications/the-llama-3-herd-of-models

Architecture description:

An auto-regressive language model that uses an optimized transformer architecture, with context length of 8K tokens. The model uses Grouped-Query Attention.

Description of the training:

Training Time is 1.3M GPU hours on hardware of type H100-80GB.

Pretrain data:

Llama 3 was pretrained on over 15 trillion tokens. The pretraining data includes data from publicly available sources and has a cutoff of March 2023.

License:

https://llama.meta.com/llama3/license/

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 - 8192

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
Meta-Llama-3-8B
MERA
0.267 0.133 0.5 0.133 0.033 0.133 0 - 0 0.033 0 0 0.1 0 0.133 0.267 0.067 0 0.033 0.033 0 0.333 0.033 0.033 0.067 0.167 0.033 0.033 0.133 0.1 0.133
Model, team Honest Helpful Harmless
Meta-Llama-3-8B
MERA
0.623 0.644 0.621
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
Meta-Llama-3-8B
MERA
0.437 0.5 0.618 0.709 0.614 0.701 0.67 0.608 0.59 0.52 0.333 0.365 0.33 0.565 0.622 0.569 0.6 0.521 0.367 0.565 0.282 0.673 0.31 0.451 0.402 0.6 0.366 0.556 0.682 0.4 0.69 0.752 0.546 0.737 0.543 0.47 0.38 0.642 0.344 0.478 0.641 0.463 0.503 0.379 0.64 0.412 0.672 0.381 0.379 0.495 0.37 0.713 0.515 0.534 0.58 0.697 0.637
Model, team SIM FL STA
Meta-Llama-3-8B
MERA
0.868 0.581 0.171
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
Meta-Llama-3-8B
MERA
0.578 0.644 0.567 0.528 0.724 0.69 0.569 0.596 0.692 0.585 0.641 0.625 0.467 0.628 0.602 0.481 0.626 0.667 0.333 0.702 0.421 0.695 0.622 0.609 0.644 0.621 0.667 0.509 0.86 0.644 0.8 0.692 0.607 0.719 0.5 0.661 0.422 0.733 0.386 0.631 0.686 0.698 0.711 0.6 0.862 0.8 0.741 0.659 0.6 0.789 0.667 0.667 0.722 0.584 0.581 0.497 0.744
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B
MERA
0.209 0.199 0.198 0.177 0.166
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B
MERA
0.17 0.146 0.171 0.176 0.128
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B
MERA
0.107 0.117 0.109 0.096 0.095
Model, team Women Men LGBT Nationalities Migrants Other
Meta-Llama-3-8B
MERA
0.704 0.629 0.529 0.649 0.286 0.623