README.md

---
language: 
- multilingual
- af
- am
- ar
- az
- be
- bg
- bn
- ca
- ceb
- co
- cs
- cy
- da
- de
- el
- en
- eo
- es
- et
- eu
- fa
- fi
- fil
- fr
- fy
- ga
- gd
- gl
- gu
- ha
- haw
- hi
- hmn
- ht
- hu
- hy
- ig
- is
- it
- iw
- ja
- jv
- ka
- kk
- km
- kn
- ko
- ku
- ky
- la
- lb
- lo
- lt
- lv
- mg
- mi
- mk
- ml
- mn
- mr
- ms
- mt
- my
- ne
- nl
- no
- ny
- pa
- pl
- ps
- pt
- ro
- ru
- sd
- si
- sk
- sl
- sm
- sn
- so
- sq
- sr
- st
- su
- sv
- sw
- ta
- te
- tg
- th
- tr
- uk
- und
- ur
- uz
- vi
- xh
- yi
- yo
- zh
- zu
license: apache-2.0
datasets:
- multi_nli
- xnli
- dbpedia_14
- SetFit/bbc-news
- squad_v2
- race
- knowledgator/events_classification_biotech
- facebook/anli
- SetFit/qnli
metrics:
- accuracy
- f1
pipeline_tag: zero-shot-classification
tags:
- classification
- information-extraction
- zero-shot
---

**comprehend-it-multilang-base**

This is an encoder-decoder model based on [mT5-base](https://huggingface.co/google/mt5-base) that was trained on multi-language natural language inference datasets as well as on multiple text classification datasets. 

The model demonstrates a better contextual understanding of text and verbalized label because both inputs are encoded by different parts of a model - encoder and decoder respectively.

The zero-shot classifier supports nearly 100 languages and can work in both directions, meaning that labels and text can belong to different languages.

#### Install the neccessary libraries before using it
Because of the different model architecture, we can't use transformers' "zero-shot-classification" pipeline. For that, we developed a special library called [LiqFit](https://github.com/Knowledgator/LiqFit/tree/main). 
If you haven't install sentencepiece library you need to install it as well to use T5 tokenizers.

```bash
pip install liqfit sentencepiece
```

#### With the LiqFit pipeline

The model can be loaded with the `zero-shot-classification` pipeline like so:

```python
from liqfit.pipeline import ZeroShotClassificationPipeline
from liqfit.models import T5ForZeroShotClassification
from transformers import T5Tokenizer

model = T5ForZeroShotClassification.from_pretrained('knowledgator/comprehend_it-multilingual-t5-base')
tokenizer = T5Tokenizer.from_pretrained('knowledgator/comprehend_it-multilingual-t5-base')
classifier = ZeroShotClassificationPipeline(model=model, tokenizer=tokenizer,
                                                      hypothesis_template = '{}', encoder_decoder = True)
```

You can then use this pipeline to classify sequences into any of the class names you specify.

```python
sequence_to_classify = "one day I will see the world"
candidate_labels = ['travel', 'cooking', 'dancing']
classifier(sequence_to_classify, candidate_labels, multi_label=False)
{'sequence': 'one day I will see the world',
 'labels': ['travel', 'cooking', 'dancing'],
 'scores': [0.7350383996963501, 0.1484801471233368, 0.1164814680814743]}
```

Amoung Enlish you can use the model for many other languages, such as Ukrainian:

```python
sequence_to_classify = "Одного дня я побачу цей світ."
candidate_labels = ['подорож', 'кулінарія', 'танці']
classifier(sequence_to_classify, candidate_labels, multi_label=False)
{'sequence': 'Одного дня я побачу цей світ.',
 'labels': ['подорож', 'кулінарія', 'танці'],
 'scores': [0.6393420696258545, 0.2657214105129242, 0.09493650496006012]}
```

The model works even if labels and text are different languages:
```python
sequence_to_classify = "Одного дня я побачу цей світ"
candidate_labels = ['travel', 'cooking', 'dancing']
classifier(sequence_to_classify, candidate_labels, multi_label=False)
{'sequence': 'Одного дня я побачу цей світ',
 'labels': ['travel', 'cooking', 'dancing'],
 'scores': [0.7676175236701965, 0.15484870970249176, 0.07753374427556992]}
```

### Benchmarking
Below, you can see the F1 score on several text classification datasets. All tested models were not fine-tuned on those datasets and were tested in a zero-shot setting.
| Model                       | IMDB | AG_NEWS | Emotions |
|-----------------------------|------|---------|----------|
| [Bart-large-mnli (407 M)](https://huggingface.co/facebook/bart-large-mnli)      | 0.89 | 0.6887  | 0.3765   |
| [Deberta-base-v3 (184 M)](https://huggingface.co/cross-encoder/nli-deberta-v3-base)      | 0.85 | 0.6455  | 0.5095   |
| [Comprehendo (184M)](https://huggingface.co/knowledgator/comprehend_it-base)           | 0.90 | 0.7982  | 0.5660   |
| [Comprehendo-multi-lang (390M)](https://huggingface.co/knowledgator/comprehend-it-multilang-base)           | 0.88 | 0.8372  | -  |
| SetFit [BAAI/bge-small-en-v1.5 (33.4M)](https://huggingface.co/BAAI/bge-small-en-v1.5) | 0.86 | 0.5636 | 0.5754 |

### Future reading
Check our blogpost - ["The new milestone in zero-shot capabilities (it’s not Generative AI)."](https://medium.com/p/9b5a081fbf27), where we highlighted possible use-cases of the model and why next-token prediction is not the only way to achive amazing zero-shot capabilites.
While most of the AI industry is focused on generative AI and decoder-based models, we are committed to developing encoder-based models.
We aim to achieve the same level of generalization for such models as their decoder brothers. Encoders have several wonderful properties, such as bidirectional attention, and they are the best choice for many information extraction tasks in terms of efficiency and controllability.

### Feedback
We value your input! Share your feedback and suggestions to help us improve our models.
Fill out the feedback [form](https://forms.gle/5CPFFuLzNWznjcpL7)

### Join Our Discord
Connect with our community on Discord for news, support, and discussion about our models.
Join [Discord](https://discord.gg/dkyeAgs9DG)