Meta-Llama-3-70B-Instruct

MERA Created at 22.09.2024 20:45
0.586
The overall result
77
Place in the rating
Weak tasks:
135
RWSD
142
PARus
48
RCB
196
ruEthics
78
MultiQ
95
ruWorldTree
65
ruOpenBookQA
102
CheGeKa
97
ruMMLU
50
ruHateSpeech
104
ruDetox
38
ruHHH
114
ruTiE
32
ruHumanEval
112
USE
109
MathLogicQA
181
ruMultiAr
121
SimpleAr
167
LCS
61
BPS
74
ruModAr
81
MaMuRAMu
57
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.132 Accuracy
RCB 0.584 / 0.491 Accuracy F1 macro
USE 0.289 Grade norm
RWSD 0.581 Accuracy
PARus 0.88 Accuracy
ruTiE 0.798 Accuracy
MultiQ 0.557 / 0.366 F1 Exact match
CheGeKa 0.324 / 0.255 F1 Exact match
ruModAr 0.733 Exact match
MaMuRAMu 0.815 Accuracy
ruMultiAr 0.298 Exact match
ruCodeEval 0.376 / 0.606 / 0.671 Pass@k
MathLogicQA 0.534 Accuracy
ruWorldTree 0.97 / 0.969 Accuracy F1 macro
ruOpenBookQA 0.92 / 0.92 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 0.99 Accuracy
ruMMLU 0.715 Accuracy
SimpleAr 0.987 Exact match
ruHumanEval 0.501 / 0.674 / 0.72 Pass@k
ruHHH 0.871
ruHateSpeech 0.849
ruDetox 0.288
ruEthics
Correct God Ethical
Virtue 0.314 0.371 0.401
Law 0.316 0.348 0.393
Moral 0.357 0.389 0.445
Justice 0.285 0.344 0.374
Utilitarianism 0.273 0.297 0.339

Information about the submission

Mera version
v.1.2.0
Torch Version
2.3.1
The version of the codebase
6ec745e0
CUDA version
12.1
Precision of the model weights
bfloat16
Seed
1234
Batch
1
Transformers version
4.43.1
The number of GPUs and their type
8 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Meta-Llama-3-70B-Instruct

Model size

70.6B

Model type:

Opened

SFT

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 6.4M 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 December 2023. The fine-tuning data includes publicly available instruction datasets, as well as over 10M human-annotated examples.

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, ruhumaneval, rucodeeval, rutie - 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-3-70B-Instruct
MERA
0.467 0.4 0.867 0.167 0.067 0.433 0.033 - 0 0 0.033 0.033 0.167 0.1 0 0.417 0.067 0 0.2 0.067 0.067 0.733 0.333 0.2 0.067 0.583 0.233 0.267 0.5 0.567 0.6
Model, team Honest Helpful Harmless
Meta-Llama-3-70B-Instruct
MERA
0.803 0.864 0.948
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-3-70B-Instruct
MERA
0.719 0.53 0.822 0.855 0.83 0.836 0.796 0.733 0.802 0.749 0.596 0.587 0.62 0.815 0.83 0.737 0.81 0.833 0.567 0.802 0.534 0.854 0.46 0.728 0.554 0.84 0.531 0.667 0.731 0.57 0.76 0.826 0.755 0.899 0.758 0.722 0.46 0.871 0.556 0.655 0.864 0.816 0.724 0.684 0.89 0.611 0.907 0.533 0.514 0.723 0.63 0.873 0.779 0.807 0.89 0.824 0.886
Model, team SIM FL STA
Meta-Llama-3-70B-Instruct
MERA
0.52 0.726 0.801
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-3-70B-Instruct
MERA
0.733 0.911 0.717 0.685 0.855 0.776 0.655 0.702 0.962 0.769 0.795 0.742 0.625 0.798 0.807 0.716 0.748 0.822 0.807 0.895 0.789 0.898 0.911 0.834 0.867 0.864 0.795 0.754 0.912 0.822 0.867 0.885 0.777 0.912 0.727 0.821 0.8 0.867 0.737 0.815 0.89 0.921 0.911 0.956 0.914 0.889 0.914 0.886 0.8 0.93 0.867 0.884 0.785 0.766 0.744 0.76 0.9
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-70B-Instruct
MERA
0.314 0.316 0.357 0.285 0.273
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-70B-Instruct
MERA
0.371 0.348 0.389 0.344 0.297
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-70B-Instruct
MERA
0.401 0.393 0.445 0.374 0.339
Model, team Women Men LGBT Nationalities Migrants Other
Meta-Llama-3-70B-Instruct
MERA
0.852 0.743 0.882 0.865 0.571 0.918