Meta-Llama-3.1-70B-Instruct

MERA Created at 22.09.2024 20:52
0.596
The overall result
67
Place in the rating
Weak tasks:
48
RWSD
63
PARus
67
RCB
35
ruEthics
49
MultiQ
84
ruWorldTree
89
ruOpenBookQA
72
CheGeKa
66
ruMMLU
54
ruHateSpeech
88
ruDetox
57
ruHHH
83
ruTiE
38
ruHumanEval
100
USE
112
MathLogicQA
168
ruMultiAr
142
SimpleAr
112
LCS
141
BPS
66
ruModAr
71
MaMuRAMu
91
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.15 Accuracy
RCB 0.578 / 0.487 Accuracy F1 macro
USE 0.309 Grade norm
RWSD 0.665 Accuracy
PARus 0.922 Accuracy
ruTiE 0.835 Accuracy
MultiQ 0.585 / 0.383 F1 Exact match
CheGeKa 0.355 / 0.284 F1 Exact match
ruModAr 0.773 Exact match
MaMuRAMu 0.824 Accuracy
ruMultiAr 0.306 Exact match
ruCodeEval 0.242 / 0.461 / 0.549 Pass@k
MathLogicQA 0.525 Accuracy
ruWorldTree 0.973 / 0.973 Accuracy F1 macro
ruOpenBookQA 0.905 / 0.905 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 0.967 Accuracy
ruMMLU 0.757 Accuracy
SimpleAr 0.983 Exact match
ruHumanEval 0.428 / 0.668 / 0.738 Pass@k
ruHHH 0.86
ruHateSpeech 0.845
ruDetox 0.303
ruEthics
Correct God Ethical
Virtue 0.454 0.415 0.51
Law 0.455 0.401 0.5
Moral 0.49 0.434 0.546
Justice 0.422 0.374 0.479
Utilitarianism 0.405 0.379 0.448

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-3.1-70B-Instruct

Model size

70.6B

Model type:

Opened

SFT

Architecture description:

A pretrained and instruction tuned 70B auto-regressive generative model with Grouped-Query Attention. Optimized for multilingual dialogue use cases.

Description of the training:

The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.

Pretrain data:

Llama 3.1 was pretrained on ~15 trillion tokens of data from publicly available sources. The fine-tuning data includes publicly available instruction datasets, as well as over 25M synthetically generated examples. The pretraining data has a cutoff of December 2023.

License:

https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/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 - 131072

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-3.1-70B-Instruct
MERA
0.567 0.467 0.8 0.2 0 0.467 0.067 - 0.033 0 0.033 0.033 0.167 0.1 0.067 0.467 0.067 0 0.033 0.067 0.133 0.733 0.333 0.2 0.133 0.642 0.267 0.4 0.567 0.5 0.567
Model, team Honest Helpful Harmless
Meta-Llama-3.1-70B-Instruct
MERA
0.803 0.831 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-3.1-70B-Instruct
MERA
0.696 0.554 0.868 0.833 0.856 0.891 0.806 0.781 0.861 0.776 0.579 0.603 0.57 0.843 0.845 0.76 0.76 0.903 0.644 0.817 0.741 0.889 0.48 0.699 0.589 0.92 0.6 0.657 0.767 0.53 0.79 0.876 0.755 0.939 0.777 0.778 0.5 0.897 0.609 0.768 0.879 0.875 0.71 0.695 0.904 0.75 0.912 0.474 0.518 0.752 0.69 0.886 0.8 0.857 0.9 0.855 0.927
Model, team SIM FL STA
Meta-Llama-3.1-70B-Instruct
MERA
0.749 0.671 0.645
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-3.1-70B-Instruct
MERA
0.756 0.911 0.767 0.685 0.868 0.845 0.724 0.754 0.942 0.723 0.885 0.75 0.633 0.814 0.848 0.765 0.729 0.8 0.789 0.86 0.895 0.898 0.8 0.852 0.844 0.803 0.808 0.754 0.895 0.867 0.867 0.846 0.804 0.912 0.727 0.875 0.889 0.844 0.754 0.8 0.89 0.889 0.889 0.933 0.931 0.933 0.879 0.909 0.723 0.947 0.889 0.913 0.823 0.753 0.721 0.772 0.9
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3.1-70B-Instruct
MERA
0.454 0.455 0.49 0.422 0.405
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3.1-70B-Instruct
MERA
0.415 0.401 0.434 0.374 0.379
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3.1-70B-Instruct
MERA
0.51 0.5 0.546 0.479 0.448
Model, team Women Men LGBT Nationalities Migrants Other
Meta-Llama-3.1-70B-Instruct
MERA
0.852 0.743 0.824 0.838 0.714 0.918