Google's UMT5 XL

MERA Created at 12.01.2024 14:48

0.201

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	0.12	Accuracy
RCB	0.326 / 0.185	Accuracy F1 macro
USE	0.001	Grade norm
RWSD	0.5	Accuracy
PARus	0.506	Accuracy
ruTiE	0.528	Accuracy
MultiQ	0.013 / 0.003	F1 Exact match
CheGeKa	0.001 / 0	F1 Exact match
ruModAr	0.0	Exact match
ruMultiAr	0.0	Exact match
MathLogicQA	0.261	Accuracy
ruWorldTree	0.269 / 0.255	Accuracy F1 macro
ruOpenBookQA	0.23 / 0.223	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

0.494

Accuracy

ruMMLU

0.254

Accuracy

SimpleAr

0.0

Exact match

ruHumanEval

0 / 0 / 0

Pass@k

ruHHH

0.489

ruHateSpeech

0.525

ruDetox

0.209

ruEthics

	Correct	God	Ethical
Virtue	-0.005	0.007	-0.034
Law	0.011	0.021	-0.058
Moral	-0.005	-0.005	-0.061
Justice	0	-0.003	-0.025
Utilitarianism	-0.031	-0.006	-0.055

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 XL

Link to the ML model:

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

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 XL MERA	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-

Model, team	Honest	Helpful	Harmless
Google's UMT5 XL MERA	0.492	0.458	0.517

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 XL MERA	0.3	0.188	0.3	0.429	0.333	0.1	0.067	0.176	0.3	0.1	0.182	0.3	0.2	0.308	0.182	0.3	0.4	0.185	0.4	0.2	0.2	0.288	0.2	0.235	0.1	0.091	0.188	0.071	0.5	0.091	0.4	0.111	0.3	0.2	0.364	0.2	0.2	0.19	0.2	0.3	0.241	0	0.5	0.5	0.375	0.2	0.3	0.6	0.1	0.4	0.273	0.25	0.294	0.333	0.292	0.212	0.222

Model, team	SIM	FL	STA
Google's UMT5 XL MERA	0.589	0.466	0.667

Coorect

Good

Ethical

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 XL MERA	-0.005	0.011	-0.005	0	-0.031

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 XL MERA	0.007	0.021	-0.005	-0.003	-0.006

Model, team	Virtue	Law	Moral	Justice	Utilitarianism
Google's UMT5 XL MERA	-0.034	-0.058	-0.061	-0.025	-0.055

Model, team	Women	Men	LGBT	Nationalities	Migrants	Other
Google's UMT5 XL MERA	0.509	0.629	0.588	0.568	0.286	0.475

Google's UMT5 XL

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