---
license: cc-by-nc-sa-4.0
datasets:
- wi_locness
- matejklemen/falko_merlin
- paws
- paws-x
- asset
language:
- en
- de
- es
- ar
- ja
- ko
- zh
metrics:
- bleu
- rouge
- sari
- accuracy
library_name: transformers
widget:
- text: >-
    Umschreiben sie den satz: When I grow up, I start to understand what he said
    is quite right.
  example_title: GEC (de|en)
- text: >-
    문장의 간단한 버전 작성: Cuando se pueden mantener tasas de flujo comparables, los
    resultados son altos.
  example_title: Simplification (ko|es)
- text: 'Paraphrase this: いちごは物語を紹介し、読者をイベントに導くと彼は言った。'
  example_title: Paraphrase (en|ja)
pipeline_tag: text2text-generation
---

# Model Card for mEdIT-xl

The `medit-xl` model was obtained by fine-tuning the `MBZUAI/bactrian-x-llama-7b-lora` model on the mEdIT dataset.

**Paper:** mEdIT: Multilingual Text Editing via Instruction Tuning

**Authors:** Vipul Raheja, Dimitris Alikaniotis, Vivek Kulkarni, Bashar Alhafni, Dhruv Kumar

## Model Details

### Model Description

- **Language(s) (NLP)**: Arabic, Chinese, English, German, Japanese, Korean, Spanish
- **Finetuned from model:** `MBZUAI/bactrian-x-llama-7b-lora`

### Model Sources

- **Repository:** https://github.com/vipulraheja/medit
- **Paper:** https://arxiv.org/abs/2402.16472v1

## How to use

Given an edit instruction and an original text, our model can generate the edited version of the text.<br>

![task_specs](https://cdn-uploads.huggingface.co/production/uploads/60985a0547dc3dbf8a976607/816ZY2t0XPCpMMd6Z072K.png)

Specifically, our models support both multi-lingual and cross-lingual text revision. Note that the input and output texts are always in the same language. The monolingual
vs. cross-lingual setting is determined by comparing the language of the edit instruction in relation to the language of the input text.

### Instruction format

Adherence to the following instruction format is essential; failure to do so may result in the model producing less-than-ideal results.

```
instruction_tokens = [
    "Instruction",
    "Anweisung",
    ...
]

input_tokens = [
    "Input",
    "Aporte",
    ...
]

output_tokens = [
    "Output",
    "Produzione",
    ...
]

task_descriptions = [
    "Fix grammatical errors in this sentence",  # <-- GEC task
    "Umschreiben Sie den Satz",                 # <-- Paraphrasing
    ...
]
```

**The entire list of possible instructions, input/output tokens, and task descriptions can be found in the Appendix of our paper.**

```
prompt_template = """### <instruction_token>:\n<task_description>\n### <input_token>:\n<input>\n### <output_token>:\n\n"""
```

Note that the tokens and the task description need not be in the language of the input (in the case of cross-lingual revision).


### Run the model

**Make sure you have the following libraries installed:**
```
- peft
- protobuf
- sentencepiece
- tokenizers
- torch
- transformers
```

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "grammarly/medit-xl"
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(model_id)

# English GEC using Japanese instructions
prompt = '### 命令:\n文章を文法的にする\n### 入力:\nI has small cat ,\n### 出力:\n\n'

inputs = tokenizer(prompt, return_tensors='pt')

outputs = model.generate(**inputs, max_new_tokens=20)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# --> I have a small cat ,

# German GEC using Japanese instructions
prompt = '### 命令:\n文章を文法的にする\n### 入力:\nIch haben eines kleines Katze ,\n### 出力:\n\n'

# ...
# --> Ich habe eine kleine Katze ,
```

#### Software
https://github.com/vipulraheja/medit

## Citation

**BibTeX:**
```
@article{raheja2023medit,
      title={mEdIT: mEdIT: Multilingual Text Editing via Instruction Tuning}, 
      author={Vipul Raheja and Dimitris Alikaniotis and Vivek Kulkarni and Bashar Alhafni and Dhruv Kumar},
      year={2024},
      eprint={2402.16472v1},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

**APA:**
Raheja, V., Alikaniotis, D., Kulkarni, V., Alhafni, B., & Kumar, D. (2024). MEdIT: Multilingual Text Editing via Instruction Tuning. ArXiv. /abs/2402.16472