Phi-3.5-MoE-instruct

MERA Created at 22.09.2024 21:16
0.501
The overall result
152
Place in the rating
Weak tasks:
320
RWSD
155
PARus
314
RCB
50
ruEthics
308
MultiQ
147
ruWorldTree
137
ruOpenBookQA
172
CheGeKa
164
ruMMLU
195
ruHateSpeech
277
ruDetox
228
ruHHH
387
ruTiE
68
ruHumanEval
186
USE
153
MathLogicQA
281
ruMultiAr
225
SimpleAr
117
LCS
83
BPS
165
ruModAr
177
MaMuRAMu
55
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.148 Accuracy
RCB 0.511 / 0.448 Accuracy F1 macro
USE 0.192 Grade norm
RWSD 0.519 Accuracy
PARus 0.876 Accuracy
ruTiE 0.558 Accuracy
MultiQ 0.334 / 0.158 F1 Exact match
CheGeKa 0.181 / 0.139 F1 Exact match
ruModAr 0.525 Exact match
MaMuRAMu 0.721 Accuracy
ruMultiAr 0.252 Exact match
ruCodeEval 0.377 / 0.583 / 0.646 Pass@k
MathLogicQA 0.474 Accuracy
ruWorldTree 0.949 / 0.948 Accuracy F1 macro
ruOpenBookQA 0.878 / 0.877 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 0.984 Accuracy
ruMMLU 0.647 Accuracy
SimpleAr 0.969 Exact match
ruHumanEval 0.295 / 0.479 / 0.537 Pass@k
ruHHH 0.702
ruHateSpeech 0.762
ruDetox 0.171
ruEthics
Correct God Ethical
Virtue 0.389 0.443 0.572
Law 0.417 0.424 0.533
Moral 0.42 0.466 0.599
Justice 0.366 0.368 0.509
Utilitarianism 0.357 0.389 0.512

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
9b26db97
CUDA version
12.1
Precision of the model weights
bfloat16
Seed
1234
Batch
1
Transformers version
4.44.2
The number of GPUs and their type
8 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Phi-3.5-MoE-instruct

Model size

41.9B

Model type:

Opened

SFT

Additional links:

https://arxiv.org/abs/2404.14219

Architecture description:

The Phi-3.5-MoE adopts an Mixture-of-Experts (MoE) architecture and belongs to the Phi-3 family.

Description of the training:

Pre-training is performed in two disjoint and sequential phases. Phase-1 comprises mostly of web sources aimed at teaching the model general knowledge and language understanding. Phase-2 merges even more heavily filtered webdata (a subset used in Phase-1) with some synthetic data that teach the model logical reasoning and various niche skills. Post-training of Phi-3 went through two stages, including supervised finetuning (SFT) and direct preference optimization (DPO).

Pretrain data:

The training data consists of heavily filtered publicly available web data (according to the “educational level”) from various open internet sources, as well as synthetic LLM-generated data.

License:

MIT

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 - 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
Phi-3.5-MoE-instruct
MERA
0.5 0.233 0.733 0.1 0.033 0.1 0.067 - 0.033 0.033 0.033 0.033 0.233 0 0.133 0.433 0 0 0.033 0.033 0.033 0.533 0.067 0.033 0.033 0.3 0.067 0.133 0.533 0.433 0.267
Model, team Honest Helpful Harmless
Phi-3.5-MoE-instruct
MERA
0.689 0.678 0.741
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
Phi-3.5-MoE-instruct
MERA
0.578 0.512 0.73 0.829 0.725 0.761 0.806 0.662 0.713 0.664 0.474 0.619 0.5 0.722 0.693 0.662 0.7 0.778 0.5 0.71 0.437 0.737 0.44 0.601 0.527 0.76 0.481 0.648 0.698 0.57 0.73 0.818 0.638 0.848 0.74 0.709 0.41 0.823 0.536 0.631 0.793 0.761 0.579 0.637 0.835 0.644 0.76 0.404 0.472 0.652 0.57 0.781 0.751 0.811 0.79 0.745 0.756
Model, team SIM FL STA
Phi-3.5-MoE-instruct
MERA
0.356 0.654 0.77
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
Phi-3.5-MoE-instruct
MERA
0.578 0.881 0.683 0.593 0.868 0.724 0.672 0.614 0.673 0.692 0.718 0.675 0.558 0.682 0.643 0.568 0.654 0.778 0.632 0.754 0.667 0.746 0.756 0.787 0.733 0.742 0.692 0.614 0.912 0.733 0.822 0.808 0.688 0.912 0.682 0.714 0.778 0.844 0.596 0.692 0.804 0.857 0.778 0.733 0.914 0.822 0.793 0.75 0.662 0.895 0.778 0.681 0.722 0.727 0.651 0.591 0.811
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Phi-3.5-MoE-instruct
MERA
0.389 0.417 0.42 0.366 0.357
Model, team Virtue Law Moral Justice Utilitarianism
Phi-3.5-MoE-instruct
MERA
0.443 0.424 0.466 0.368 0.389
Model, team Virtue Law Moral Justice Utilitarianism
Phi-3.5-MoE-instruct
MERA
0.572 0.533 0.599 0.509 0.512
Model, team Women Men LGBT Nationalities Migrants Other
Phi-3.5-MoE-instruct
MERA
0.796 0.714 0.765 0.622 0.429 0.852