GigaChat-3.1-Ultra

GigaChat Создан 28.03.2026 08:28
0.361
Общий результат

Оценки по задачам лидерборда

Таблица скроллится влево

Задача Результат Метрика
YABLoCo 0.024 / 0.019
EM pass@k
stRuCom 0.296
chrF
RealCode 0.349 / 0.971
pass@k execution_success
UnitTests 0.245
CodeBLEU
ruCodeEval 0.542 / 0.612 / 0.628
pass@k
JavaTestGen 0.159 / 0.449
pass@k compile@1
ruHumanEval 0.539 / 0.578 / 0.585
pass@k
RealCodeJava 0.289 / 0.96
pass@k execution_success
CodeLinterEval 0.489 / 0.507 / 0.518
pass@k
ruCodeReviewer 0.015 / 0.131 / 0.065 / 0.073 / 0.075
chrF BLEU judge@1 judge@5 judge@10
CodeCorrectness 0.728
EM

Информация о сабмите

Версия MERA
v1.0.0
Версия Torch
2.9.0
Версия кодовой базы
0ac3a14
Версия CUDA
12.8
Precision весов модели
auto
Сид
1234
Батч
1
Версия transformers
4.57.1
Количество GPU и их тип
1 x NVIDIA A100-SXM4-80GB
Архитектура
gigachat-completion

Команда:

GigaChat

Название ML-модели:

GigaChat-3.1-Ultra

Ссылка на ML-модель:

https://huggingface.co/ai-sage/GigaChat3.1-702B-A36B

Размер модели

715.0B

Тип модели:

Открытая

SFT

Описание архитектуры:

GigaChat 3.1 Ultra is the flagship instruct model of the GigaChat family. It is a large-scale Mixture-of-Experts (MoE) model with 702B total parameters and 36B active parameters, designed for multilingual assistant workloads, reasoning, code, tool use, and large-cluster deployment.

Описание обучения:

The model underwent Pretraining, Stage-1.5, SFT and DPO stages.

Данные претрейна:

The base GigaChat 3 training corpus spans 10 languages and includes books, academic material, code datasets, and mathematics datasets. All data goes through deduplication, language filtering, and automatic quality checks based on heuristics and classifiers.

Лицензия:

MIT

Параметры инференса

Параметры генерации:
realcode - do_sample=true;max_gen_toks=4096;temperature=0.7;repetition_penalty=1.05;top_p=0.8;until=["<|endoftext|>","<|im_end|>"]; \nrealcodejava - do_sample=true;max_gen_toks=4096;temperature=0.7;repetition_penalty=1.05;top_p=0.8;until=["<|endoftext|>","<|im_end|>"]; \njavatestgen - do_sample=true;max_gen_toks=4096;temperature=0.2;top_p=0.9;until=["<|endoftext|>","<|im_end|>"]; \nyabloco_oracle - max_gen_toks=2048;do_sample=false;until=["<|endoftext|>","<|im_end|>","\n\n\n","\\sclass\\s","\\sdef\\s","^def\\s","^class\\s","^if\\s","@","^#"]; \nunittests - do_sample=false;max_gen_toks=1024;until=["\n\n"]; \ncodecorrectness - until=["\n\n"];do_sample=false;temperature=0; \ncodelintereval - do_sample=true;temperature=0.6;max_gen_toks=1024;until=["\n\n"]; \nrucodereviewer - temperature=0;do_sample=false;max_gen_toks=1000;until=["\n\n"]; \nstrucom - do_sample=false;max_gen_toks=512;until=["\n\n"]; \nrucodeeval_code - do_sample=true;temperature=0.6;max_gen_toks=1024;until=["\nclass","\ndef","\n#","\nif","\nprint"]; \nruhumaneval_code - do_sample=true;temperature=0.6;max_gen_toks=1024;until=["\nclass","\ndef","\n#","\nif","\nprint"];