Llama-3.2-3B-Instruct

MERA Created at 04.11.2024 17:38
0.38
The overall result
359
Place in the rating
Weak tasks:
388
RWSD
429
PARus
480
RCB
408
ruEthics
392
MultiQ
372
ruWorldTree
393
ruOpenBookQA
432
CheGeKa
369
ruMMLU
435
ruHateSpeech
396
ruDetox
318
ruHHH
416
ruTiE
96
ruHumanEval
377
USE
443
MathLogicQA
465
ruMultiAr
382
SimpleAr
201
LCS
453
BPS
216
ruModAr
357
MaMuRAMu
119
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.126 Accuracy
RCB 0.409 / 0.355 Accuracy F1 macro
USE 0.086 Grade norm
RWSD 0.504 Accuracy
PARus 0.606 Accuracy
ruTiE 0.544 Accuracy
MultiQ 0.269 / 0.176 F1 Exact match
CheGeKa 0.025 / 0.019 F1 Exact match
ruModAr 0.497 Exact match
MaMuRAMu 0.54 Accuracy
ruMultiAr 0.113 Exact match
ruCodeEval 0.171 / 0.42 / 0.512 Pass@k
MathLogicQA 0.318 Accuracy
ruWorldTree 0.741 / 0.741 Accuracy F1 macro
ruOpenBookQA 0.633 / 0.632 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 0.663 Accuracy
ruMMLU 0.453 Accuracy
SimpleAr 0.875 Exact match
ruHumanEval 0.21 / 0.442 / 0.524 Pass@k
ruHHH 0.618
ruHateSpeech 0.551
ruDetox 0.118
ruEthics
Correct God Ethical
Virtue 0.103 0.167 0.184
Law 0.098 0.164 0.162
Moral 0.103 0.181 0.174
Justice 0.103 0.141 0.152
Utilitarianism 0.105 0.152 0.138

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
1 x NVIDIA H100 80GB HBM3
Architecture
vllm

Team:

MERA

Name of the ML model:

Llama-3.2-3B-Instruct

Model size

3.0B

Model type:

Opened

SFT

Architecture description:

Llama 3.2 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.

Description of the training:

-

Pretrain data:

Llama 3.2 was pretrained on up to 9 trillion tokens of data from publicly available sources. For the 1B and 3B Llama 3.2 models, we incorporated logits from the Llama 3.1 8B and 70B models into the pretraining stage of the model development, where outputs (logits) from these larger models were used as token-level targets. Knowledge distillation was used after pruning to recover performance. In post-training we used a similar recipe as Llama 3.1 and produced final chat models by doing several rounds of alignment on top of the pre-trained model. Each round involved Supervised Fine-Tuning (SFT), Rejection Sampling (RS), and Direct Preference Optimization (DPO).

License:

llama3.2

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

System prompt:
Реши задачу по инструкции ниже. Не давай никаких объяснений и пояснений к своему ответу. Не пиши ничего лишнего. Пиши только то, что указано в инструкции. Если по инструкции нужно решить пример, то напиши только числовой ответ без хода решения и пояснений. Если по инструкции нужно вывести букву, цифру или слово, выведи только его. Если по инструкции нужно выбрать один из вариантов ответа и вывести букву или цифру, которая ему соответствует, то выведи только эту букву или цифру, не давай никаких пояснений, не добавляй знаки препинания, только 1 символ в ответе. Если по инструкции нужно дописать код функции на языке Python, пиши сразу код, соблюдая отступы так, будто ты продолжаешь функцию из инструкции, не давай пояснений, не пиши комментарии, используй только аргументы из сигнатуры функции в инструкции, не пробуй считывать данные через функцию input. Не извиняйся, не строй диалог. Выдавай только ответ и ничего больше.

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
Llama-3.2-3B-Instruct
MERA
0.167 0.033 0.5 0.033 0 0.1 0 - 0.033 0 0.067 0.033 0 0 0.067 0.217 0.067 0.033 0 0 0.033 0.1 0.067 0.033 0.033 0.167 0 0.1 0.133 0.133 0.067
Model, team Honest Helpful Harmless
Llama-3.2-3B-Instruct
MERA
0.672 0.593 0.586
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
Llama-3.2-3B-Instruct
MERA
0.4 0.404 0.467 0.585 0.51 0.582 0.553 0.518 0.451 0.43 0.307 0.349 0.34 0.546 0.507 0.526 0.45 0.375 0.333 0.534 0.306 0.509 0.26 0.457 0.375 0.57 0.382 0.407 0.641 0.4 0.61 0.736 0.466 0.707 0.536 0.406 0.35 0.503 0.391 0.478 0.561 0.412 0.51 0.401 0.517 0.407 0.559 0.319 0.344 0.398 0.41 0.633 0.421 0.462 0.54 0.636 0.487
Model, team SIM FL STA
Llama-3.2-3B-Instruct
MERA
0.726 0.533 0.373
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
Llama-3.2-3B-Instruct
MERA
0.489 0.485 0.417 0.509 0.5 0.5 0.431 0.596 0.423 0.523 0.667 0.542 0.392 0.628 0.398 0.383 0.589 0.533 0.368 0.649 0.386 0.576 0.489 0.515 0.644 0.545 0.513 0.491 0.754 0.622 0.733 0.603 0.518 0.561 0.545 0.5 0.444 0.711 0.456 0.554 0.604 0.46 0.667 0.489 0.776 0.689 0.655 0.614 0.6 0.702 0.6 0.536 0.684 0.519 0.419 0.409 0.644
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
Llama-3.2-3B-Instruct
MERA
0.103 0.098 0.103 0.103 0.105
Model, team Virtue Law Moral Justice Utilitarianism
Llama-3.2-3B-Instruct
MERA
0.167 0.164 0.181 0.141 0.152
Model, team Virtue Law Moral Justice Utilitarianism
Llama-3.2-3B-Instruct
MERA
0.184 0.162 0.174 0.152 0.138
Model, team Women Men LGBT Nationalities Migrants Other
Llama-3.2-3B-Instruct
MERA
0.574 0.4 0.647 0.568 0.429 0.574