README.md

---
language:
- ru
tags:
- spellchecking
- M2M100
- pytorch
- natural language generation
license: mit
datasets:
- ai-forever/spellcheck_benchmark
metrics:
- precision
- recall
- f1
library_name: transformers
model-index:
- name: sage-mt5-large
  results:
  - task:
      type: text-generation
    dataset:
      type: spellcheck_benchmark
      name: RUSpellRU 
    metrics:
    - name: Precision
      type: precision
      value: 88.8
      verified: false
    - name: Recall
      type: recall
      value: 71.5
      verified: false
    - name: F1 
      type: f1
      value: 79.2
      verified: false
  - task:
      type: text-generation
    dataset:
      type: spellcheck_benchmark
      name: MultidomainGold 
    metrics:
    - name: Precision
      type: precision
      value: 63.8
      verified: false
    - name: Recall
      type: recall
      value: 61.1
      verified: false
    - name: F1 
      type: f1
      value: 62.4
      verified: false
  - task:
      type: text-generation
    dataset:
      type: spellcheck_benchmark
      name: MedSpellchecker
    metrics:
    - name: Precision
      type: precision
      value: 78.8
      verified: false
    - name: Recall
      type: recall
      value: 71.4
      verified: false
    - name: F1 
      type: f1
      value: 74.9
      verified: false
  - task:
      type: text-generation
    dataset:
      type: spellcheck_benchmark
      name: GitHubTypoCorpusRu
    metrics:
    - name: Precision
      type: precision
      value: 47.1
      verified: false
    - name: Recall
      type: recall
      value: 42.9
      verified: false
    - name: F1 
      type: f1
      value: 44.9
      verified: false
---
# sage-m2m100-1.2B model

![banner](images/sage_banner.jpg)

## Summary

The model corrects spelling errors and typos by bringing all the words in the text to the norm of the Russian language.
Corrector was trained based on the model [M2M100-1.2B](https://huggingface.co/facebook/m2m100_1.2B). 
An extensive dataset with “artificial” errors was taken as a training corpus: the corpus was assembled on the basis of the Russian-language Wikipedia and transcripts of Russian-language videos, then typos and spelling errors were automatically introduced into it using the library [SAGE](https://github.com/ai-forever/sage).
The model is the fine-tuned version of the [pre-train](https://huggingface.co/ai-forever/RuM2M100-1.2B). 

## Public references
- [SAGE library announcement](https://youtu.be/yFfkV0Qjuu0), DataFest 2023
- [Paper about synthetic error generation methods](https://www.dialog-21.ru/media/5914/martynovnplusetal056.pdf), Dialogue 2023
- [SAGE EACL 2024 paper](https://aclanthology.org/2024.findings-eacl.10/)


## Examples
| Input | Output |
| --- | --- |
| Думю ешцъа лет череа 10 ретроспективно просматривотьэ то будкетцц мне невероя тна ин те р но | Думаю что лет через 10 ретроспективно просматривать это будет мне невероятно интересно |
| Основая цель мероприятия - практическая отработка навыков по оказанию помощи гражданам, попавшим в ДТП, а также повышение и совершенствование уровня профессиональной подготовки сотрудников МЧС при проведении аварийно-спасательных работ по ликвидации последствий дорожно-транспортных проишествий, сокращение временных показателей реагирования. | Основная цель мероприятия - практическая отработка навыков по оказанию помощи гражданам, попавшим в ДТП, а также повышение и совершенствование уровня профессиональной подготовки сотрудников МЧС при проведении аварийно-спасательных работ по ликвидации последствий дорожно-транспортных происшествий, сокращение временных показателей реагирования. |
| прийдя в МГТУ я был удивлен никого необноружив там… | придя в МГТУ я был удивлен никого не обнаружив там |
|  |  |

## Metrics
### Quality
Below are automatic metrics for determining the correctness of the spell checkers. 
We compare our solution with both open automatic spell checkers and the ChatGPT family of models on all four available datasets:
- **RUSpellRU**: texts collected from ([LiveJournal](https://www.livejournal.com/media)), with manually corrected typos and errors;
- **MultidomainGold**: examples from 7 text sources, including the open web, news, social media, reviews, subtitles, policy documents and literary works;
- **MedSpellChecker**: texts with errors from medical anamnesis;
- **GitHubTypoCorpusRu**: spelling errors and typos in commits from [GitHub](https://github.com);

**RUSpellRU**
| Model | Precision | Recall | F1 |
| --- | --- | --- | --- |
| sage-m2m100-1.2B | 88.8 | 71.5 | 79.2 |
| sage-ai-service | 93.5 | 82.4 | 87.6 | 
| gpt-3.5-turbo | 39.6 | 62.3 | 48.5 | 
| gpt-4 | 69.5 | 81.0 | 74.8 | 
| Yandex.Speller | 83.0 | 59.8 | 69.5 |
| JamSpell | 42.1 | 32.8 | 36.9 |
| HunSpell | 31.3 | 34.9 | 33.0 |

**MultidomainGold**
| Model | Precision | Recall | F1 |
| --- | --- | --- | --- |
| sage-m2m100-1.2B | 63.8 | 61.1 | 62.4 |
| sage-ai-service | 70.9 | 68.8 | 69.9 | 
| gpt-3.5-turbo | 17.8 | 56.1 | 27.0 | 
| gpt-4 | 31.1 | 78.1 | 44.5 |
| Yandex.Speller | 52.9 | 51.4 | 52.2 |
| JamSpell | 25.7 | 30.6 | 28.0 |
| HunSpell | 16.2 | 40.1 | 23.0 |

**MedSpellChecker**
| Model | Precision | Recall | F1 |
| --- | --- | --- | --- |
| sage-m2m100-1.2B | 78.8 | 71.4 | 74.9 |
| sage-ai-service | 73.4 | 76.2 | 74.9 | 
| gpt-3.5-turbo | 15.1 | 53.6 | 23.5 | 
| gpt-4 | 48.9 | 88.7 | 63.1 | 
| Yandex.Speller | 80.6 | 47.8 | 60.0 |
| JamSpell | 24.6 | 29.7 | 26.9 |
| HunSpell | 10.3 | 40.2 | 16.4 |

**GitHubTypoCorpusRu**
| Model | Precision | Recall | F1 |
| --- | --- | --- | --- |
| sage-m2m100-1.2B | 47.1 | 42.9 | 44.9 |
| sage-ai-service | 76.1 | 51.2 | 61.2 | 
| gpt-3.5-turbo | 23.7 | 43.9 | 30.8 | 
| gpt-4 | 34.7  | 60.5 | 44.1| 
| Yandex.Speller | 67.7 | 37.5 | 48.3 |
| JamSpell | 49.5 | 29.9 | 37.3 |
| HunSpell | 28.5 | 30.7 | 29.6 |

## How to use
```python
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer

path_to_model = "ai-forever/sage-m2m100-1.2B"
model = M2M100ForConditionalGeneration.from_pretrained(path_to_model)
tokenizer = M2M100Tokenizer.from_pretrained(path_to_model, src_lang="ru", tgt_lang="ru")

sentence = "прийдя в МГТУ я был удивлен никого необноружив там…"
encodings = tokenizer(sentence, return_tensors="pt")
generated_tokens = model.generate(
        **encodings, forced_bos_token_id=tokenizer.get_lang_id("ru"))
answer = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)

print(answer)
#["прийдя в МГТУ я был удивлен никого не обнаружив там..."]
```

## Resources
- [SAGE library](https://github.com/ai-forever/sage), GitHub
- [sage-fredt5-large](https://huggingface.co/ai-forever/sage-fredt5-large), HuggingFace
- [sage-fredt5-distilled-95m](https://huggingface.co/ai-forever/sage-fredt5-distilled-95m), HuggingFace
- [sage-m2m100-1.2B](https://huggingface.co/ai-forever/sage-m2m100-1.2B), HuggingFace
- [sage-mt5-large](https://huggingface.co/ai-forever/sage-mt5-large), HuggingFace

## License
Model [M2M100-1.2B](https://huggingface.co/facebook/m2m100_1.2B), on the basis of which our solution is made, and its source code are supplied under the MIT open license. 
Our solution also comes with MIT license.

## Specifications
- File size: 5 Gb;
- Framework: pytorch
- Format: AI Service
- Version: v2.0
- Developer: SberDevices, AGI NLP

## Contacts
[email protected]