Meta-Llama-3-8B-Instruct

MERA Created at 22.09.2024 20:48
0.477
The overall result
185
Place in the rating
Weak tasks:
149
RWSD
254
PARus
263
RCB
242
ruEthics
332
MultiQ
297
ruWorldTree
269
ruOpenBookQA
229
CheGeKa
316
ruMMLU
304
ruHateSpeech
167
ruDetox
234
ruHHH
220
ruTiE
55
ruHumanEval
303
USE
370
MathLogicQA
329
ruMultiAr
284
SimpleAr
258
LCS
187
BPS
124
ruModAr
280
MaMuRAMu
46
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.112 Accuracy
RCB 0.527 / 0.44 Accuracy F1 macro
USE 0.125 Grade norm
RWSD 0.577 Accuracy
PARus 0.808 Accuracy
ruTiE 0.695 Accuracy
MultiQ 0.321 / 0.16 F1 Exact match
CheGeKa 0.136 / 0.099 F1 Exact match
ruModAr 0.642 Exact match
MaMuRAMu 0.628 Accuracy
ruMultiAr 0.227 Exact match
ruCodeEval 0.396 / 0.615 / 0.665 Pass@k
MathLogicQA 0.356 Accuracy
ruWorldTree 0.829 / 0.828 Accuracy F1 macro
ruOpenBookQA 0.758 / 0.755 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 0.958 Accuracy
ruMMLU 0.506 Accuracy
SimpleAr 0.94 Exact match
ruHumanEval 0.364 / 0.553 / 0.61 Pass@k
ruHHH 0.697
ruHateSpeech 0.664
ruDetox 0.224
ruEthics
Correct God Ethical
Virtue 0.261 0.303 0.302
Law 0.26 0.284 0.276
Moral 0.294 0.319 0.316
Justice 0.229 0.259 0.264
Utilitarianism 0.25 0.284 0.273

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-Instruct

Model size

8.0B

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 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. 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, 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-Instruct
MERA
0.133 0.267 0.567 0.067 0 0.167 0.033 - 0 0.033 0.033 0.033 0.033 0 0.167 0.417 0 0.033 0 0.1 0 0.2 0.067 0.033 0.067 0.092 0.033 0.067 0.433 0.233 0.233
Model, team Honest Helpful Harmless
Meta-Llama-3-8B-Instruct
MERA
0.738 0.644 0.707
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-Instruct
MERA
0.43 0.452 0.553 0.697 0.621 0.627 0.641 0.579 0.577 0.525 0.325 0.389 0.36 0.611 0.61 0.578 0.64 0.521 0.344 0.573 0.249 0.678 0.37 0.491 0.384 0.64 0.38 0.565 0.669 0.41 0.61 0.736 0.503 0.717 0.574 0.491 0.32 0.635 0.318 0.409 0.662 0.467 0.552 0.374 0.64 0.426 0.637 0.359 0.376 0.508 0.38 0.717 0.505 0.534 0.68 0.715 0.596
Model, team SIM FL STA
Meta-Llama-3-8B-Instruct
MERA
0.506 0.673 0.731
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-Instruct
MERA
0.467 0.644 0.6 0.602 0.539 0.638 0.569 0.526 0.692 0.569 0.603 0.642 0.45 0.69 0.573 0.58 0.701 0.644 0.421 0.702 0.298 0.763 0.511 0.598 0.644 0.636 0.641 0.596 0.86 0.689 0.756 0.692 0.661 0.719 0.515 0.518 0.556 0.689 0.456 0.677 0.722 0.587 0.778 0.667 0.879 0.8 0.707 0.682 0.569 0.772 0.689 0.667 0.696 0.597 0.535 0.52 0.733
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B-Instruct
MERA
0.261 0.26 0.294 0.229 0.25
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B-Instruct
MERA
0.303 0.284 0.319 0.259 0.284
Model, team Virtue Law Moral Justice Utilitarianism
Meta-Llama-3-8B-Instruct
MERA
0.302 0.276 0.316 0.264 0.273
Model, team Women Men LGBT Nationalities Migrants Other
Meta-Llama-3-8B-Instruct
MERA
0.704 0.371 0.765 0.703 0.429 0.738