Qwen2-0.5B-Instruct

MERA Created at 22.09.2024 21:42
0.255
The overall result
534
Place in the rating
Weak tasks:
445
RWSD
545
PARus
560
RCB
543
MultiQ
535
ruWorldTree
556
ruOpenBookQA
578
CheGeKa
538
ruMMLU
549
ruHateSpeech
552
ruDetox
450
ruHHH
504
ruTiE
230
ruHumanEval
552
USE
479
MathLogicQA
479
ruMultiAr
455
SimpleAr
450
LCS
557
BPS
432
ruModAr
543
MaMuRAMu
238
ruCodeEval
+18
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.078 Accuracy
RCB 0.345 / 0.287 Accuracy F1 macro
USE 0.04 Grade norm
RWSD 0.492 Accuracy
PARus 0.498 Accuracy
ruTiE 0.506 Accuracy
MultiQ 0.144 / 0.061 F1 Exact match
CheGeKa 0.006 / 0 F1 Exact match
ruModAr 0.243 Exact match
MaMuRAMu 0.337 Accuracy
ruMultiAr 0.099 Exact match
ruCodeEval 0.007 / 0.023 / 0.037 Pass@k
MathLogicQA 0.307 Accuracy
ruWorldTree 0.425 / 0.414 Accuracy F1 macro
ruOpenBookQA 0.37 / 0.353 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 0.497 Accuracy
ruMMLU 0.32 Accuracy
SimpleAr 0.655 Exact match
ruHumanEval 0.03 / 0.084 / 0.116 Pass@k
ruHHH 0.517
ruHateSpeech 0.487
ruDetox 0.055
ruEthics
Correct God Ethical
Virtue 0.043 0.035 -0.021
Law 0.062 0.051 0.029
Moral 0.057 0.059 -0.001
Justice 0.063 0.041 -0.008
Utilitarianism 0.025 0.066 -0.032

Information about the submission

Mera version
v.1.2.0
Torch Version
2.4.0
The version of the codebase
6ec745e0
CUDA version
12.1
Precision of the model weights
float16
Seed
1234
Batch
1
Transformers version
4.43.2
The number of GPUs and their type
1 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

MERA

Name of the ML model:

Qwen2-0.5B-Instruct

Model size

494.0M

Model type:

Opened

SFT

Additional links:

https://arxiv.org/pdf/2407.10671

Architecture description:

Qwen2-0.5B-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
Qwen2-0.5B-Instruct
MERA
0.033 0.033 0.3 0.1 0 0.267 0 - 0 0 0 0 0 0 0.033 0.15 0.033 0 0.067 0 0 0 0 0 0 0.008 0 0.033 0.067 0 0.067
Model, team Honest Helpful Harmless
Qwen2-0.5B-Instruct
MERA
0.525 0.559 0.466
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
Qwen2-0.5B-Instruct
MERA
0.393 0.295 0.355 0.38 0.35 0.368 0.379 0.334 0.37 0.336 0.219 0.262 0.37 0.491 0.364 0.361 0.33 0.264 0.144 0.26 0.289 0.263 0.33 0.341 0.205 0.28 0.284 0.426 0.269 0.31 0.33 0.471 0.374 0.434 0.317 0.278 0.26 0.368 0.219 0.32 0.318 0.254 0.317 0.3 0.343 0.319 0.333 0.248 0.323 0.316 0.36 0.35 0.29 0.273 0.37 0.479 0.337
Model, team SIM FL STA
Qwen2-0.5B-Instruct
MERA
0.546 0.525 0.347
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
Qwen2-0.5B-Instruct
MERA
0.222 0.317 0.317 0.417 0.289 0.293 0.345 0.263 0.327 0.415 0.372 0.3 0.192 0.403 0.228 0.296 0.402 0.422 0.298 0.333 0.246 0.186 0.356 0.32 0.378 0.379 0.449 0.228 0.649 0.289 0.556 0.321 0.429 0.333 0.288 0.321 0.311 0.311 0.368 0.338 0.29 0.302 0.333 0.289 0.586 0.467 0.31 0.295 0.477 0.368 0.333 0.333 0.418 0.325 0.395 0.263 0.367
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-0.5B-Instruct
MERA
0.043 0.062 0.057 0.063 0.025
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-0.5B-Instruct
MERA
0.035 0.051 0.059 0.041 0.066
Model, team Virtue Law Moral Justice Utilitarianism
Qwen2-0.5B-Instruct
MERA
-0.021 0.029 -0.001 -0.008 -0.032
Model, team Women Men LGBT Nationalities Migrants Other
Qwen2-0.5B-Instruct
MERA
0.435 0.514 0.412 0.595 0.143 0.557