Ryoo72/llava-pretrain-vsft

theblackcat102/llava-pretrain reformated for VSFT with TRL's SFT Trainer.
Referenced the format of HuggingFaceH4/llava-instruct-mix-vsft

I used this code for format changes.
Warning: I'm not completely sure if I used this code. Please use it only as a reference.

import json
from datasets import load_dataset
from huggingface_hub import login

login(token="YOUR TOKEN")

def transform_format(example):
    conversations = json.loads(example['conversations'])
    
    messages = []
    for conv in conversations:
        if conv['from'] == 'human':
            value = conv['value']
            # 이미지 태그가 중간에 있는지 확인
            if '<image>' in value and not (value.startswith('<image>') or value.endswith('<image>')):
                raise ValueError(f"Image tag found in the middle of text: {value}")
            
            text = value.replace('<image>\n', '').replace('\n<image>', '')
            content = []
            
            # 이미지가 텍스트 앞에 있는 경우
            if value.startswith('<image>'):
                content = [
                    {"index": 0, "text": None, "type": "image"},
                    {"index": None, "text": text, "type": "text"}
                ]
            # 이미지가 텍스트 뒤에 있는 경우
            elif value.endswith('<image>'):
                content = [
                    {"index": None, "text": text, "type": "text"},
                    {"index": 0, "text": None, "type": "image"}
                ]
            else:
                # 이미지 태그가 없는 경우
                content = [
                    {"index": None, "text": text, "type": "text"}
                ]
                
            messages.append({
                "content": content,
                "role": "user"
            })
        else:  # 'gpt' case
            messages.append({
                "content": [
                    {"index": None, "text": conv['value'], "type": "text"}
                ],
                "role": "assistant"
            })
    
    return {
        'messages': messages,
        'images': example['image']
    }

dataset = load_dataset("theblackcat102/llava-pretrain")

try:
    transformed_dataset = dataset.map(
        transform_format,
        num_proc=64,
        batch_size=128,
        remove_columns=dataset['train'].column_names,
        desc="Transforming dataset"
    )
except ValueError as e:
    print(f"Error during transformation: {e}")

transformed_dataset.push_to_hub(
    "Ryoo72/llava-pretrain-vsft",
    private=False
)