T-lite-0.1

Т-Банк Created at 20.09.2024 15:54
0.51
The overall result
137
Place in the rating
Weak tasks:
85
RWSD
249
PARus
226
RCB
350
ruEthics
126
MultiQ
241
ruWorldTree
246
ruOpenBookQA
125
CheGeKa
264
ruMMLU
237
ruHateSpeech
251
ruDetox
335
ruHHH
231
ruTiE
80
ruHumanEval
314
USE
325
MathLogicQA
275
ruMultiAr
288
SimpleAr
116
LCS
444
BPS
110
ruModAr
232
MaMuRAMu
89
ruCodeEval
+19
Hide

Ratings for leaderboard tasks

The table will scroll to the left

Task name Result Metric
LCS 0.148 Accuracy
RCB 0.537 / 0.448 Accuracy F1 macro
USE 0.122 Grade norm
RWSD 0.612 Accuracy
PARus 0.81 Accuracy
ruTiE 0.688 Accuracy
MultiQ 0.482 / 0.346 F1 Exact match
CheGeKa 0.305 / 0.248 F1 Exact match
ruModAr 0.669 Exact match
MaMuRAMu 0.689 Accuracy
ruMultiAr 0.253 Exact match
ruCodeEval 0.252 / 0.479 / 0.573 Pass@k
MathLogicQA 0.375 Accuracy
ruWorldTree 0.888 / 0.887 Accuracy F1 macro
ruOpenBookQA 0.785 / 0.785 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 0.689 Accuracy
ruMMLU 0.549 Accuracy
SimpleAr 0.939 Exact match
ruHumanEval 0.266 / 0.497 / 0.579 Pass@k
ruHHH 0.601
ruHateSpeech 0.732
ruDetox 0.18
ruEthics
Correct God Ethical
Virtue 0.203 0.208 0.272
Law 0.232 0.259 0.275
Moral 0.239 0.24 0.3
Justice 0.189 0.207 0.227
Utilitarianism 0.146 0.214 0.254

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
bfloat16
Seed
1234
Batch
1
Transformers version
4.43.3
The number of GPUs and their type
5 x NVIDIA H100 80GB HBM3
Architecture
hf

Team:

Т-Банк

Name of the ML model:

T-lite-0.1

Model size

8.0B

Model type:

Opened

Pretrain

Architecture description:

T-lite is a continual pretraining model designed specifically for the Russian language, enabling the creation of large language model applications in Russian. This model aims to improve the quality of Russian text generation and provide domain-specific and cultural knowledge relevant to the Russian context. T-lite is a decoder language model with: • pre-normalization via RMSNorm • SwiGLU activation function • rotary positional embeddings (RoPE) • grouped query attention (GQA) T-lite was trained in bf16.

Description of the training:

Training Hyperparameters Authors employed the Decoupled AdamW optimizer with β1 = 0.9, β2 = 0.95, and eps = 1.0e-8. The learning rate was set to 1.0e-5 with a constant schedule and a warmup period of 10 steps during stage 1, and a cosine schedule during stage 2. Weight decay was applied at a rate of 1.0e-6, and gradient clipping was performed with a maximum norm of 1.0. The maximum sequence length was set to 8192. Each batch contained approximately 6 million tokens. Hardware Configuration & Performance Training was conducted on 96 A100 GPUs with 80GB memory each, using Fully Sharded Data Parallel (FSDP) with full shard/hybrid shard strategies. The setup achieved a throughput of 3000 tokens/sec/GPU, with 100B tokens being processed in approximately 4 days. We achieved a 0.59 Model FLOPs Utilization (MFU).

Pretrain data:

Stage 1. Massive continual pre-training • 300B tokens * 0.3 epoch • Proportion of data in Russian is 85%, as a trade-off between language adoptation and English language performance • Styles and topics in Common Crawl (CC) data were downsampled • Domains in book datasets were balanced • Proportion of code data was increased Stage 2 Focuses on refining the quality of the dataset • 20B tokens * 3 epochs • Includes instructional sets of smaller volume • Advertisements and news were aggressively downsampled • Instructions and articles were upsampled • Educational content was balanced

License:

https://huggingface.co/AnatoliiPotapov/T-lite-0.1/blob/main/LICENSE

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 - 8192

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
T-lite-0.1
Т-Банк
0.367 0.2 0.533 0.133 0.033 0.367 0.067 - 0 0 0 0.033 0.133 0 0 0.15 0 0 0.067 0.033 0.033 0.4 0.067 0.033 0.067 0.133 0.1 0.2 0.1 0.133 0.2
Model, team Honest Helpful Harmless
T-lite-0.1
Т-Банк
0.557 0.593 0.655
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
T-lite-0.1
Т-Банк
0.593 0.506 0.625 0.765 0.621 0.756 0.68 0.627 0.623 0.596 0.289 0.357 0.34 0.657 0.708 0.546 0.65 0.632 0.367 0.672 0.321 0.713 0.35 0.503 0.402 0.62 0.401 0.556 0.682 0.45 0.67 0.719 0.613 0.848 0.6 0.534 0.34 0.723 0.364 0.433 0.727 0.574 0.545 0.424 0.715 0.421 0.725 0.311 0.379 0.52 0.4 0.759 0.582 0.563 0.62 0.691 0.731
Model, team SIM FL STA
T-lite-0.1
Т-Банк
0.774 0.666 0.393
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
T-lite-0.1
Т-Банк
0.622 0.782 0.65 0.602 0.75 0.672 0.5 0.684 0.827 0.662 0.692 0.65 0.533 0.721 0.661 0.617 0.738 0.711 0.509 0.719 0.404 0.814 0.622 0.716 0.667 0.742 0.705 0.719 0.912 0.689 0.822 0.744 0.625 0.789 0.621 0.768 0.556 0.778 0.474 0.631 0.767 0.73 0.778 0.6 0.897 0.778 0.862 0.659 0.754 0.877 0.711 0.71 0.734 0.623 0.419 0.573 0.744
Coorect
Good
Ethical
Model, team Virtue Law Moral Justice Utilitarianism
T-lite-0.1
Т-Банк
0.203 0.232 0.239 0.189 0.146
Model, team Virtue Law Moral Justice Utilitarianism
T-lite-0.1
Т-Банк
0.208 0.259 0.24 0.207 0.214
Model, team Virtue Law Moral Justice Utilitarianism
T-lite-0.1
Т-Банк
0.272 0.275 0.3 0.227 0.254
Model, team Women Men LGBT Nationalities Migrants Other
T-lite-0.1
Т-Банк
0.704 0.657 0.765 0.703 0.571 0.852