README.md

---
language:
- ko  # Example: fr
license: apache-2.0  # Example: apache-2.0 or any license from https://hf.co/docs/hub/repositories-licenses
library_name: transformers  # Optional. Example: keras or any library from https://github.com/huggingface/hub-docs/blob/main/js/src/lib/interfaces/Libraries.ts
tags:
- text2text-generation  # Example: audio
datasets:
- aihub  # Example: common_voice. Use dataset id from https://hf.co/datasets
metrics:
- bleu  # Example: wer. Use metric id from https://hf.co/metrics
- rouge

# Optional. Add this if you want to encode your eval results in a structured way.
model-index:
- name: ko-barTNumText
  results:
  - task:
      type: text2text-generation             # Required. Example: automatic-speech-recognition
      name: text2text-generation             # Optional. Example: Speech Recognition
    metrics:
      - type: bleu         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: 0.9161441917016176       # Required. Example: 20.90
        name: eval_bleu         # Optional. Example: Test WER
        verified: true              # Optional. If true, indicates that evaluation was generated by Hugging Face (vs. self-reported).
      - type: rouge1         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: 0.9502159661745533       # Required. Example: 20.90
        name: eval_rouge1         # Optional. Example: Test WER
        verified: true              # Optional. If true, indicates that evaluation was generated by Hugging Face (vs. self-reported).
      - type: rouge2         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: 0.9313935147887745       # Required. Example: 20.90
        name: eval_rouge2       # Optional. Example: Test WER
        verified: true              # Optional. If true, indicates that evaluation was generated by Hugging Face (vs. self-reported).
      - type: rougeL         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: 0.950015374196916       # Required. Example: 20.90
        name: eval_rougeL        # Optional. Example: Test WER
        verified: true              # Optional. If true, indicates that evaluation was generated by Hugging Face (vs. self-reported).
      - type: rougeLsum         # Required. Example: wer. Use metric id from https://hf.co/metrics
        value: 0.9500390902948073       # Required. Example: 20.90
        name: eval_rougeLsum        # Optional. Example: Test WER
        verified: true              # Optional. If true, indicates that evaluation was generated by Hugging Face (vs. self-reported).
---

# ko-barTNumText(TNT Model🧨): Try Number To Korean Reading(숫자를 한글로 바꾸는 모델)

## Table of Contents
- [ko-barTNumText(TNT Model🧨): Try Number To Korean Reading(숫자를 한글로 바꾸는 모델)](#ko-bartnumtexttnt-model-try-number-to-korean-reading숫자를-한글로-바꾸는-모델)
  - [Table of Contents](#table-of-contents)
  - [Model Details](#model-details)
  - [Uses](#uses)
  - [Evaluation](#evaluation)
  - [How to Get Started With the Model](#how-to-get-started-with-the-model)


## Model Details
- **Model Description:**
뭔가 찾아봐도 모델이나 알고리즘이 딱히 없어서 만들어본 모델입니다. <br />
BartForConditionalGeneration Fine-Tuning Model For Number To Korean <br />
BartForConditionalGeneration으로 파인튜닝한, 숫자를 한글로 변환하는 Task 입니다. <br />

- Dataset use [Korea aihub](https://aihub.or.kr/aihubdata/data/list.do?currMenu=115&topMenu=100&srchDataRealmCode=REALM002&srchDataTy=DATA004) <br />
I can't open my fine-tuning datasets for my private issue <br />
데이터셋은 Korea aihub에서 받아서 사용하였으며, 파인튜닝에 사용된 모든 데이터를 사정상 공개해드릴 수는 없습니다. <br />

- Korea aihub data is ONLY permit to Korean!!!!!!! <br />
aihub에서 데이터를 받으실 분은 한국인일 것이므로, 한글로만 작성합니다. <br />
정확히는 음성전사를 철자전사로 번역하는 형태로 학습된 모델입니다. (ETRI 전사기준) <br />

- In case, ten million, some people use 10 million or some people use 10000000, so this model is crucial for training datasets <br />
천만을 1000만 혹은 10000000으로 쓸 수도 있기에, Training Datasets에 따라 결과는 상이할 수 있습니다. <br />
**수관형사와 수 의존명사의 띄어쓰기에 따라 결과가 확연히 달라질 수 있습니다. (쉰살, 쉰 살 -> 쉰살, 50살)** https://eretz2.tistory.com/34 <br />
일단은 기준을 잡고 치우치게 학습시키기엔 어떻게 사용될지 몰라, 학습 데이터 분포에 맡기도록 했습니다. (쉰 살이 더 많을까 쉰살이 더 많을까!?)
- **Developed by:**  Yoo SungHyun(https://github.com/YooSungHyun)
- **Language(s):** Korean
- **License:** apache-2.0
- **Parent Model:** See the [kobart-base-v2](https://huggingface.co/gogamza/kobart-base-v2) for more information about the pre-trained base model.
  
  
## Uses
This Model is inferenced token BACKWARD. so, you have to `flip` before `tokenizer.decode()` <br />
해당 모델은 inference시 역순으로 예측합니다. (밥을 6시에 먹었어 -> 어 먹었 시에 여섯 을 밥) <br />
때문에 `tokenizer.decode`를 수행하기 전에, `flip`으로 역순으로 치환해주세요.

Want see more detail follow this URL [KoGPT_num_converter](https://github.com/ddobokki/KoGPT_num_converter) <br /> and see `bart_inference.py` and `bart_train.py`
```python
class BartText2TextGenerationPipeline(Text2TextGenerationPipeline):
    def postprocess(self, model_outputs, return_type=ReturnType.TEXT, clean_up_tokenization_spaces=False):
        records = []
        reversed_model_outputs = torch.flip(model_outputs["output_ids"][0], dims=[-1])
        for output_ids in reversed_model_outputs:
            if return_type == ReturnType.TENSORS:
                record = {f"{self.return_name}_token_ids": output_ids}
            elif return_type == ReturnType.TEXT:
                record = {
                    f"{self.return_name}_text": self.tokenizer.decode(
                        output_ids,
                        skip_special_tokens=True,
                        clean_up_tokenization_spaces=clean_up_tokenization_spaces,
                    )
                }
            records.append(record)
        return records
```
## Evaluation
Just using `evaluate-metric/bleu` and `evaluate-metric/rouge` in huggingface `evaluate` library <br />
[Training wanDB URL](https://wandb.ai/bart_tadev/BartForConditionalGeneration/runs/2dt1d2b0?workspace=user-bart_tadev)
## How to Get Started With the Model
```python
from transformers.pipelines import Text2TextGenerationPipeline
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
texts = ["그러게 누가 6시까지 술을 마시래?"]
tokenizer = AutoTokenizer.from_pretrained(
    args.model_name_or_path,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
    args.model_name_or_path,
)
# BartText2TextGenerationPipeline is implemented above (see 'Use')
seq2seqlm_pipeline = BartText2TextGenerationPipeline(model=model, tokenizer=tokenizer)
kwargs = {
    "min_length": args.min_length,
    "max_length": args.max_length,
    "num_beams": args.beam_width,
    "do_sample": args.do_sample,
    "num_beam_groups": args.num_beam_groups,
}
pred = seq2seqlm_pipeline(texts, **kwargs)
print(pred)
# 그러게 누가 여섯 시까지 술을 마시래?
```