---
datasets:
- kresnik/zeroth_korean
metrics:
- bleu
- cer
base_model:
- microsoft/Phi-4-multimodal-instruct
model-index:
- name: Phi-4-mm-inst-zeroth-kor
  results:
  - task:
      type: speech-to-text-translation
    dataset:
      type: seastar105/fleurs_ko_en_test
      name: fleurs (ko-en test intersection)
    metrics:
    - type: bleu
      name: ko2en
      value: 7.07
    - type: bleu
      name: ko2en-cot
      value: 9.19
    - type: bleu
      name: en2ko (ko-mecab)
      value: 13.08
    - type: bleu
      name: en2ko-cot (ko-mecab)
      value: 9.35
  - task:
      type: automatic-speech-recognition
    dataset:
      type: kresnik/zeroth_korean
      name: zeroth_korean test
    metrics:
    - type: cer
      name: test CER
      value: 7.02
language:
- ko
---
This model is fine-tuned from [microsoft/Phi-4-multimodal-instruct](https://huggingface.co/microsoft/Phi-4-multimodal-instruct) on [kresnik/zeroth_korean](https://huggingface.co/datasets/kresnik/zeroth_korean) dataset only 1 epoch.

script for fine-tuning is [here](https://gist.github.com/seastar105/d1d8983b27611370528e3b194dcc5577#file-main-py), adapted from phi-4 repository example

model is trained only 174 steps on zeroth train set, and main purpose is to check if only korean ASR training can expand to other speech tasks(e.g. speech-to-text-translation)

## Evaluation

ASR on zeroth-test set and Speech translation on fleurs ko <-> en speech translation result. script is [here](https://gist.github.com/seastar105/d1d8983b27611370528e3b194dcc5577#file-evaluate-py), and used 1 A40.


|  Model   | zeroth-test | fleurs-ko2en | fleurs-ko2en-cot | fleurs-en2ko | fleurs-en2ko-cot |
|----------|------------|--------------|------------------|--------------|------------------|
| original |   195.92   |     5.62     |       2.45       |     6.87     |       4.35       |
| finetune (this model) |    7.02    |     7.07     |       9.19       |    13.08     |       9.35       |


## Example script

```python
orig_model_path = "microsoft/Phi-4-multimodal-instruct"
ft_model_path = "seastar105/Phi-4-mm-inst-zeroth-kor"
generation_config = GenerationConfig.from_pretrained(orig_model_path, 'generation_config.json')
processor = AutoProcessor.from_pretrained(orig_model_path, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    ft_model_path,
    trust_remote_code=True,
    torch_dtype='auto',
    _attn_implementation='flash_attention_2',
).cuda()

user_prompt = '<|user|>'
assistant_prompt = '<|assistant|>'
prompt_suffix = '<|end|>'

# task prompt is from technical report
asr_prompt = f'{user_prompt}<|audio_1|>Transcribe the audio clip into text.{prompt_suffix}{assistant_prompt}'
ast_ko_prompt = f'{user_prompt}<|audio_1|>Translate the audio to Korean.{prompt_suffix}{assistant_prompt}'
ast_cot_ko_prompt = f'{user_prompt}<|audio_1|>Transcribe the audio to text, and then translate the audio to Korean. Use <sep> as a separator between the original transcript and the translation.{prompt_suffix}{assistant_prompt}'
ast_en_prompt = f'{user_prompt}<|audio_1|>Translate the audio to English.{prompt_suffix}{assistant_prompt}'
ast_cot_en_prompt = f'{user_prompt}<|audio_1|>Transcribe the audio to text, and then translate the audio to English. Use <sep> as a separator between the original transcript and the translation.{prompt_suffix}{assistant_prompt}'

asr_ds = load_dataset("kresnik/zeroth_korean", split="test")
ast_ds = load_dataset("seastar105/fleurs_ko_en_test", split="train")

# ASR
item = asr_ds[0]
audio = (item["audio"]["array"], item["audio"]["sampling_rate"])
inputs = processor(text=asr_prompt, audios=[audio], return_tensors='pt').to(model.device)
generate_ids = model.generate(
    **inputs,
    max_new_tokens=max_new_tokens,
    generation_config=generation_config,
)
generate_ids = generate_ids[:, inputs['input_ids'].shape[1] :]
response = processor.batch_decode(
    generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0] # "몬토 킬은 자녀들이 사랑을 제대로 못 받고 크면 매우 심각한 결과가 초래된다는 결론을 내렸습니다"

# AST, EN -> KO
item = ast_ds[-1]
audio = (item["en_audio"]["array"], item["en_audio"]["sampling_rate"])
inputs = processor(text=ast_en, audios=[audio], return_tensors='pt').to(model.device)
generate_ids = model.generate(
    **inputs,
    max_new_tokens=max_new_tokens,
    generation_config=generation_config,
)
generate_ids = generate_ids[:, inputs['input_ids'].shape[1] :]
response = processor.batch_decode(
    generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0] # "가장 쉽게 접근 가능한 식물 자원은 잎과 légumes에서 접근 가능한 단백질이었을 것이다가요 하지만 이것들은 고형상 동물처럼 우리에게 소화하기 어렵습니다만 그것들이 끓여 있다면요"
```