Google's UMT5 Base

MERA Created at 13.01.2024 00:02

0.195

The overall result

The submission does not contain all the required tasks

Ratings for leaderboard tasks

The table will scroll to the left

Task name	Result	Metric
LCS	0.106	Accuracy
RCB	0.336 / 0.306	Accuracy F1 macro
USE	0	Grade norm
RWSD	0.5	Accuracy
PARus	0.468	Accuracy
ruTiE	0.526	Accuracy
MultiQ	0.002 / 0	F1 Exact match
CheGeKa	0.001 / 0	F1 Exact match
ruModAr	0.0	Exact match
ruMultiAr	0.0	Exact match
MathLogicQA	0.252	Accuracy
ruWorldTree	0.238 / 0.147	Accuracy F1 macro
ruOpenBookQA	0.243 / 0.148	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

0.523

Accuracy

ruMMLU

0.231

Accuracy

SimpleAr

0.0

Exact match

ruHumanEval

0 / 0 / 0

Pass@k

ruHHH

0.528

ruHateSpeech

0.521

ruDetox

0.005

ruEthics

	Correct	God	Ethical
Virtue	0.052	0.075	0.019
Law	0.049	0.087	0.014
Moral	0.034	0.058	0.001
Justice	0.041	0.065	0.006
Utilitarianism	0.054	0.054	0.035

Information about the submission

Mera version

Torch Version

The version of the codebase

CUDA version

Precision of the model weights

Seed

Batch

Transformers version

The number of GPUs and their type

Architecture

Team:

MERA

Name of the ML model:

Google's UMT5 Base

Link to the ML model:

https://huggingface.co/google/umt5-base

Additional links:

https://openreview.net/forum?id=kXwdL1cWOAi

Architecture description:

Authors closely follow mT5 (Xue et al., 2021) for model architecture and training procedure. Specifically, thet use an encoder-decoder Transformer architecture and the span corruption pretraining objective from T5 (Raffel et al., 2020) on a multilingual corpus consisting of 101 languages plus 6 Latin-script variants (e.g. ru-Latn). They use batch size of 1024 sequences where each sequence is defined by selecting a chunk of 568 tokens from the training corpus. This is then split into 512 input and 114 target tokens. The number of training steps is 250,000.

Description of the training:

The model architectures used in this study are the same as mT5 models, except that relative position embeddings are not shared across layers. The vocabulary size is 256,000 subwords, and byte-level fallback is enabled, so unknown tokens are broken down into UTF-8 bytes. Authors use the T5X library (Roberts et al., 2022) to train the models using Google Cloud TPUs. For pretraining, they use Adafactor optimizer (Shazeer & Stern, 2018) with a constant learning rate of 0.01 in the first 10,000 steps and inverse square root decay afterwards. For finetuning, they use Adafactor with a constant learning rate of 5e−5.

Pretrain data:

UMT5 is pretrained on the an updated version of mC4 corpus, covering 107 languages.

Training Details:

Unlike mT5, authors do not use loss normalization factor. Instead they use the number of real target tokens as the effective loss normalization. Finally, they do not factorize the second moment of the Adafactor states and use momentum, neither of which are used in T5 and mT5 studies.

License:

Apache 2.0

Strategy, generation and parameters:

Code version v.1.1.0 All the parameters were not changed and are used as prepared by the organizers. Details: - 1 x NVIDIA A100 - dtype auto - Pytorch 2.1.2 + CUDA 12.1 - Transformers 4.36.2 - Context length 512

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
Google's UMT5 Base MERA	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-

Model, team	Honest	Helpful	Harmless
Google's UMT5 Base MERA	0.459	0.559	0.569

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
Google's UMT5 Base MERA	0.3	0.313	0.2	0.257	0.238	0	0.533	0.176	0.4	0.2	0.091	0.1	0.1	0.038	0.227	0.1	0.2	0.407	0.3	0.3	0.1	0.308	0.4	0.235	0.2	0.273	0.125	0.071	0.1	0.182	0.3	0.111	0.4	0.6	0	0.1	0.2	0.429	0.3	0	0.241	0.4	0.2	0.2	0.125	0.1	0.3	0	0.3	0.2	0.364	0.063	0.176	0.467	0.25	0.273	0.111

Model, team	SIM	FL	STA
Google's UMT5 Base MERA	0.178	0.352	0.077

Coorect

Good

Ethical

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 Base MERA	0.052	0.049	0.034	0.041	0.054

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 Base MERA	0.075	0.087	0.058	0.065	0.054

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 Base MERA	0.019	0.014	0.001	0.006	0.035

Model, team	Women	Men	LGBT	Nationalities	Migrants	Other
Google's UMT5 Base MERA	0.565	0.429	0.412	0.595	0.286	0.508

Google's UMT5 Base

Ratings for leaderboard tasks

Evaluation on open tasks:

Information about the submission

Team:

Name of the ML model:

Link to the ML model:

Additional links:

Architecture description:

Description of the training:

Pretrain data:

Training Details:

License:

Strategy, generation and parameters:

Ratings by subcategory

Confirm the deletion of the sub