Qwen2-7B-Instruct

MERA Created at 22.09.2024 22:00
0.475
The overall result
189
Place in the rating
Weak tasks:
215
RWSD
224
PARus
221
RCB
185
ruEthics
286
MultiQ
207
ruWorldTree
230
ruOpenBookQA
301
CheGeKa
240
ruMMLU
188
ruHateSpeech
236
ruDetox
177
ruHHH
166
ruTiE
107
ruHumanEval
195
USE
216
MathLogicQA
232
ruMultiAr
133
SimpleAr
171
LCS
251
BPS
286
ruModAr
242
MaMuRAMu
147
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.132 Accuracy
RCB 0.539 / 0.472 Accuracy F1 macro
USE 0.183 Grade norm
RWSD 0.554 Accuracy
PARus 0.832 Accuracy
ruTiE 0.741 Accuracy
MultiQ 0.356 / 0.199 F1 Exact match
CheGeKa 0.088 / 0.053 F1 Exact match
ruModAr 0.463 Exact match
MaMuRAMu 0.687 Accuracy
ruMultiAr 0.275 Exact match
ruCodeEval 0.099 / 0.264 / 0.341 Pass@k
MathLogicQA 0.435 Accuracy
ruWorldTree 0.926 / 0.926 Accuracy F1 macro
ruOpenBookQA 0.808 / 0.807 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 0.938 Accuracy
ruMMLU 0.577 Accuracy
SimpleAr 0.985 Exact match
ruHumanEval 0.188 / 0.362 / 0.439 Pass@k
ruHHH 0.753
ruHateSpeech 0.766
ruDetox 0.187
ruEthics
Correct God Ethical
Virtue 0.386 0.33 0.374
Law 0.398 0.344 0.36
Moral 0.42 0.349 0.388
Justice 0.33 0.28 0.338
Utilitarianism 0.307 0.289 0.329

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:

Qwen2-7B-Instruct

Model size

7.6B

Model type:

Opened

SFT

Additional links:

https://arxiv.org/pdf/2407.10671

Architecture description:

Qwen2-7B-instruct is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, RoPE and RMSNorm

Description of the training:

SFT and RLHF on the mix of manually and synthetically annotated samples after pertaining with next-token prediction task.

Pretrain data:

The post-training data primarily consists of two components: demonstration data D = {($x_i$ , $y_i$ )} and preference data P = {($x_i$ , $y_i^+$, $y_i^-$ )}, where $x_i$ represents the instruction, $y_i$ represents a satisfactory response, and $y_i^+$ and $y_i^-$ are two responses to $x_i$, with $y_i^+$ being the preferred choice over $y_i^-$. The set D is utilized in SFT, whereas P is employed in RLHF. We have assembled an extensive instruction dataset featuring more than 500,000 examples that cover skills such as instruction following, coding, mathematics, logical reasoning, role-playing, multilingualism, and safety.

License:

Apache-2.0

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 - 32768

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
Qwen2-7B-Instruct
MERA
0.3 0.4 0.767 0.267 0.067 0.333 0.033 - 0 0.033 0.033 0 0.1 0.033 0.1 0.267 0.033 0 0 0.033 0 0.4 0.167 0.1 0 0.417 0.033 0 0.267 0.167 0.367
Model, team Honest Helpful Harmless
Qwen2-7B-Instruct
MERA
0.672 0.814 0.776
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
Qwen2-7B-Instruct
MERA
0.541 0.476 0.651 0.791 0.65 0.662 0.689 0.678 0.608 0.65 0.412 0.452 0.4 0.657 0.699 0.659 0.68 0.632 0.3 0.595 0.282 0.743 0.48 0.549 0.42 0.61 0.431 0.657 0.702 0.37 0.68 0.744 0.583 0.747 0.623 0.603 0.42 0.697 0.411 0.557 0.712 0.57 0.545 0.562 0.753 0.546 0.716 0.441 0.401 0.564 0.57 0.764 0.626 0.651 0.71 0.745 0.72
Model, team SIM FL STA
Qwen2-7B-Instruct
MERA
0.363 0.709 0.772
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
Qwen2-7B-Instruct
MERA
0.556 0.792 0.65 0.657 0.75 0.707 0.517 0.667 0.635 0.585 0.679 0.675 0.417 0.705 0.626 0.667 0.692 0.733 0.491 0.754 0.544 0.678 0.756 0.716 0.667 0.667 0.718 0.649 0.877 0.667 0.844 0.808 0.616 0.772 0.636 0.625 0.711 0.844 0.561 0.738 0.759 0.651 0.756 0.733 0.862 0.844 0.828 0.841 0.769 0.86 0.667 0.623 0.81 0.584 0.581 0.573 0.767
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-7B-Instruct
MERA
0.386 0.398 0.42 0.33 0.307
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-7B-Instruct
MERA
0.33 0.344 0.349 0.28 0.289
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-7B-Instruct
MERA
0.374 0.36 0.388 0.338 0.329
Model, team Women Men LGBT Nationalities Migrants Other
Qwen2-7B-Instruct
MERA
0.759 0.686 0.824 0.703 0.571 0.869