Nuextract 2.0 4B

🚀 NuExtract 2.0 4B by NuMind

NuExtract 2.0は、構造化情報抽出タスク用に特別に訓練されたモデルファミリーです。マルチモーダル入力をサポートし、多言語対応です。

API / Platform   |   Blog   |   Discord

✨ 主な機能

• 構造化情報抽出タスクに特化した訓練
• マルチモーダル入力（テキストと画像）のサポート
• 多言語対応

📦 インストール

このREADMEには具体的なインストール手順が記載されていないため、このセクションをスキップします。

💻 使用例

基本的な使用法

import torch
from transformers import AutoProcessor, AutoModelForVision2Seq

model_name = "numind/NuExtract-2.0-2B"
# model_name = "numind/NuExtract-2.0-8B"

model = AutoModelForVision2Seq.from_pretrained(model_name, 
                                               trust_remote_code=True, 
                                               torch_dtype=torch.bfloat16,
                                               attn_implementation="flash_attention_2",
                                               device_map="auto")
processor = AutoProcessor.from_pretrained(model_name, 
                                          trust_remote_code=True, 
                                          padding_side='left',
                                          use_fast=True)

# You can set min_pixels and max_pixels according to your needs, such as a token range of 256-1280, to balance performance and cost.
# min_pixels = 256*28*28
# max_pixels = 1280*28*28
# processor = AutoProcessor.from_pretrained(model_name, min_pixels=min_pixels, max_pixels=max_pixels)

高度な使用法

テキストドキュメントからの名前抽出

template = """{"names": ["string"]}"""
document = "John went to the restaurant with Mary. James went to the cinema."

# prepare the user message content
messages = [{"role": "user", "content": document}]
text = processor.tokenizer.apply_chat_template(
    messages,
    template=template, # template is specified here
    tokenize=False,
    add_generation_prompt=True,
)

print(text)
""""<|im_start|>user
# Template:
{"names": ["string"]}
# Context:
John went to the restaurant with Mary. James went to the cinema.<|im_end|> 
<|im_start|>assistant"""

image_inputs = process_all_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

# we choose greedy sampling here, which works well for most information extraction tasks
generation_config = {"do_sample": False, "num_beams": 1, "max_new_tokens": 2048}

# Inference: Generation of the output
generated_ids = model.generate(
    **inputs,
    **generation_config
)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)

print(output_text)
# ['{"names": ["John", "Mary", "James"]}']

コンテキスト内の例を使用する場合

template = """{"names": ["string"]}"""
document = "John went to the restaurant with Mary. James went to the cinema."
examples = [
    {
        "input": "Stephen is the manager at Susan's store.",
        "output": """{"names": ["-STEPHEN-", "-SUSAN-"]}"""
    }
]

messages = [{"role": "user", "content": document}]
text = processor.tokenizer.apply_chat_template(
    messages,
    template=template,
    examples=examples, # examples provided here
    tokenize=False,
    add_generation_prompt=True,
)

image_inputs = process_all_vision_info(messages, examples)
inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

# we choose greedy sampling here, which works well for most information extraction tasks
generation_config = {"do_sample": False, "num_beams": 1, "max_new_tokens": 2048}

# Inference: Generation of the output
generated_ids = model.generate(
    **inputs,
    **generation_config
)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
# ['{"names": ["-JOHN-", "-MARY-", "-JAMES-"]}']

画像入力を使用する場合

template = """{"store": "verbatim-string"}"""
document = {"type": "image", "image": "file://1.jpg"}

messages = [{"role": "user", "content": [document]}]
text = processor.tokenizer.apply_chat_template(
    messages,
    template=template,
    tokenize=False,
    add_generation_prompt=True,
)

image_inputs = process_all_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generation_config = {"do_sample": False, "num_beams": 1, "max_new_tokens": 2048}

# Inference: Generation of the output
generated_ids = model.generate(
    **inputs,
    **generation_config
)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
# ['{"store": "Trader Joe\'s"}']

バッチ推論

inputs = [
    # image input with no ICL examples
    {
        "document": {"type": "image", "image": "file://0.jpg"},
        "template": """{"store_name": "verbatim-string"}""",
    },
    # image input with 1 ICL example
    {
        "document": {"type": "image", "image": "file://0.jpg"},
        "template": """{"store_name": "verbatim-string"}""",
        "examples": [
            {
                "input": {"type": "image", "image": "file://1.jpg"},
                "output": """{"store_name": "Trader Joe's"}""",
            }
        ],
    },
    # text input with no ICL examples
    {
        "document": {"type": "text", "text": "John went to the restaurant with Mary. James went to the cinema."},
        "template": """{"names": ["string"]}""",
    },
    # text input with ICL example
    {
        "document": {"type": "text", "text": "John went to the restaurant with Mary. James went to the cinema."},
        "template": """{"names": ["string"]}""",
        "examples": [
            {
                "input": "Stephen is the manager at Susan's store.",
                "output": """{"names": ["STEPHEN", "SUSAN"]}"""
            }
        ],
    },
]

# messages should be a list of lists for batch processing
messages = [
    [
        {
            "role": "user",
            "content": [x['document']],
        }
    ]
    for x in inputs
]

# apply chat template to each example individually
texts = [
    processor.tokenizer.apply_chat_template(
        messages[i],  # Now this is a list containing one message
        template=x['template'],
        examples=x.get('examples', None),
        tokenize=False, 
        add_generation_prompt=True)
    for i, x in enumerate(inputs)
]

image_inputs = process_all_vision_info(messages, [x.get('examples') for x in inputs])
inputs = processor(
    text=texts,
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generation_config = {"do_sample": False, "num_beams": 1, "max_new_tokens": 2048}

# Batch Inference
generated_ids = model.generate(**inputs, **generation_config)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_texts = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
for y in output_texts:
    print(y)
# {"store_name": "WAL-MART"}
# {"store_name": "Walmart"}
# {"names": ["John", "Mary", "James"]}
# {"names": ["JOHN", "MARY", "JAMES"]}

テンプレート生成

xml_template = """<SportResult>
    <Date></Date>
    <Sport></Sport>
    <Venue></Venue>
    <HomeTeam></HomeTeam>
    <AwayTeam></AwayTeam>
    <HomeScore></HomeScore>
    <AwayScore></AwayScore>
    <TopScorer></TopScorer>
</SportResult>"""

messages = [
        {
            "role": "user",
            "content": [{"type": "text", "text": xml_template}],
        }
    ]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
)

image_inputs = process_all_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(
    **inputs,
    **generation_config
)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)

print(output_text[0])
# {
#     "Date": "date-time",
#     "Sport": "verbatim-string",
#     "Venue": "verbatim-string",
#     "HomeTeam": "verbatim-string",
#     "AwayTeam": "verbatim-string",
#     "HomeScore": "integer",
#     "AwayScore": "integer",
#     "TopScorer": "verbatim-string"
# }

description = "I would like to extract important details from the contract."

messages = [
        {
            "role": "user",
            "content": [{"type": "text", "text": description}],
        }
    ]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
)

image_inputs = process_all_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

generated_ids = model.generate(
    **inputs,
    **generation_config
)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)

print(output_text[0])
# {
#     "Contract": {
#         "Title": "verbatim-string",
#         "Description": "verbatim-string",
#         "Terms": [
#             {
#                 "Term": "verbatim-string",
#                 "Description": "verbatim-string"
#             }
#         ],
#         "Date": "date-time",
#         "Signatory": "verbatim-string"
#     }
# }

📚 ドキュメント

入力テンプレートのサポートタイプ

• verbatim-string - 入力にそのまま存在するテキストを抽出するようにモデルに指示します。
• string - 言い換えや抽象化を含むことができる汎用的な文字列フィールドです。
• integer - 整数です。
• number - 整数または小数です。
• date-time - ISO形式の日付です。
• 上記のいずれかの型の配列（例：["string"]）
• enum - 可能な回答のセットからの選択肢（テンプレートではオプションの配列として表されます。例：["yes", "no", "maybe"]）。
• multi-label - 複数の可能な回答を持つことができるenum（テンプレートでは二重にラップされた配列として表されます。例：[["A", "B", "C"]]）。

サンプルテンプレート

{
  "first_name": "verbatim-string",
  "last_name": "verbatim-string",
  "description": "string",
  "age": "integer",
  "gpa": "number",
  "birth_date": "date-time",
  "nationality": ["France", "England", "Japan", "USA", "China"],
  "languages_spoken": [["English", "French", "Japanese", "Mandarin", "Spanish"]]
}

サンプル出力

{
  "first_name": "Susan",
  "last_name": "Smith",
  "description": "A student studying computer science.",
  "age": 20,
  "gpa": 3.7,
  "birth_date": "2005-03-01",
  "nationality": "England",
  "languages_spoken": ["English", "French"]
}

🔧 技術詳細

このREADMEには具体的な技術詳細が記載されていないため、このセクションをスキップします。

📄 ライセンス

このプロジェクトはMITライセンスの下で提供されています。

ファインチューニング

ファインチューニングのチュートリアルノートブックは、GitHubリポジトリのcookbooksフォルダにあります。

vLLMデプロイメント

OpenAI互換APIを提供するには、以下のコマンドを実行します。

vllm serve numind/NuExtract-2.0-8B --trust_remote_code --limit-mm-per-prompt image=6 --chat-template-content-format openai

メモリの問題が発生した場合は、--max-model-lenを適切に設定してください。

モデルにリクエストを送信するには、以下のようにします。

import json
from openai import OpenAI

openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="numind/NuExtract-2.0-8B",
    temperature=0,
    messages=[
        {
            "role": "user", 
            "content": [{"type": "text", "text": "Yesterday I went shopping at Bunnings"}],
        },
    ],
    extra_body={
        "chat_template_kwargs": {
            "template": json.dumps(json.loads("""{\"store\": \"verbatim-string\"}"""), indent=4)
        },
    }
)
print("Chat response:", chat_response)

画像入力の場合は、以下のようにリクエストを構築します。"content"内の画像をプロンプト内に表示される順序で並べるようにしてください（つまり、メイン入力の前にコンテキスト内の例を配置します）。

import base64

def encode_image(image_path):
    """
    Encode the image file to base64 string
    """
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode('utf-8')

base64_image = encode_image("0.jpg")
base64_image2 = encode_image("1.jpg")

chat_response = client.chat.completions.create(
    model="numind/NuExtract-2.0-8B",
    temperature=0,
    messages=[
        {
            "role": "user", 
            "content": [
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}, # first ICL example image
            ]
        },
    ],
    extra_body={
        "chat_template_kwargs": {
            "template": json.dumps(json.loads("""{\"store\": \"verbatim-string\"}"""), indent=4)
        },
    }
)
print("Chat response:", chat_response)

⚠️ 重要提示

NuExtractを使用する際には、温度を0またはそれに非常に近い値に設定することをお勧めします。Ollamaなどの一部の推論フレームワークは、デフォルトで0.7を使用しますが、これは多くの抽出タスクに適していません。