Spaces:

numind
/

NuExtract-2.0

Sleeping

App Files Files Community

Alexandre-Numind commited on 5 days ago

Commit

3b9a7b5

verified ·

1 Parent(s): 183dac2

Upload app.py

Browse files

Files changed (1) hide show

app.py +324 -0

app.py ADDED Viewed

	@@ -0,0 +1,324 @@

+import json, torch, gradio as gr
+from transformers import AutoProcessor, AutoModelForVision2Seq
+from utils import process_all_vision_info
+model_name   = "numind/NuExtract-2.0-4B"
+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,
+)
+def run_model(image, text, template):
+    """Return (extracted_json, template_update, warning_update)."""
+    try:
+        json.loads(template)
+        template_valid = True
+    except Exception:
+        template_valid = False
+    messages = (
+        [{"role": "user", "content": text}]
+        if text
+        else [{"role": "user", "content": [{"type": "image", "image": image}]}]
+    )
+    template_arg = template if template_valid else None
+    if not template_valid:
+        messages = [{"role": "user", "content": template}]
+    chat_txt = processor.tokenizer.apply_chat_template(
+        messages, template=template_arg, tokenize=False, add_generation_prompt=True
+    )
+    img_inputs = process_all_vision_info(messages)
+    inputs = processor(
+        text=[chat_txt],
+        images=img_inputs,
+        padding=True,
+        return_tensors="pt"
+    ).to("cuda")
+    seq_len = inputs.input_ids.shape[1]
+    if seq_len > 10_000:
+        return (
+            "",
+            gr.update(),
+            gr.update(
+                value=(
+                    f"❌ **Input too long**: {seq_len} tokens "
+                    f"(limit = 10 000). Please shorten the text or image context."
+                ),
+                visible=True,
+            ),
+        )
+    ids  = model.generate(**inputs, do_sample=False, num_beams=1, max_new_tokens=4000)
+    ids2 = [o[len(i):] for i, o in zip(inputs.input_ids, ids)]
+    out  = processor.batch_decode(
+        ids2, skip_special_tokens=True, clean_up_tokenization_spaces=False
+    )[0]
+    if template_valid:
+        extract_out = json.dumps(json.loads(out), indent=4,ensure_ascii = False)
+        templ_upd   = gr.update()
+        warn_upd    = gr.update(visible=False)
+    else:
+        extract_out = ""
+        templ_upd   = gr.update(value=out)
+        warn_upd    = gr.update(
+            value="⚠️ Template wasn’t valid JSON. "
+                  "I generated one for you — check it, then press **Run** again.",
+            visible=True,
+        )
+    return extract_out, templ_upd, warn_upd
+with gr.Blocks(title="NuExtract – zero-shot structured extraction") as demo:
+    # 🚀 Banner
+    gr.HTML("""<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="UTF-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+  <title>NuExtract-2 Overview</title>
+  <style>
+    img   { display:block; margin-bottom:1rem; }
+    ul    { margin:1rem 0; padding-left:1.5rem; }
+    a     { color:#4187f6; text-decoration:none; }
+    a:hover { text-decoration:underline; }
+    h1,h2 { margin:0 0 .5rem 0; font-weight:600; }
+    pre   { overflow-x:auto; border-radius:6px; padding:1rem; }
+    code  { border-radius:4px; padding:.2em .4em; font-family:monospace; }
+    html[data-theme="dark"],
+    @media (prefers-color-scheme: dark) {
+      body { background-color:#1e1e1e;}
+      code { background-color: #2d2d2d;}
+      pre  { background-color:#2a2a2a;}
+    }
+    html[data-theme="light"],
+    @media (prefers-color-scheme: light) {
+      body { background-color#ffffff;}
+      code { background-color#f5f5f5;}
+      pre  { background-color#f5f5f5;}
+    }
+</style>
+</head>
+<body>
+<p align="center">
+    <a href="https://nuextract.ai/">
+        <img src="https://cdn.prod.website-files.com/638364a4e52e440048a9529c/64188f405afcf42d0b85b926_logo_numind_final.png"
+             alt="NuMind Logo" style="width: 200px; height: 50px;" />
+    </a>
+</p>
+<p align="center">
+    🖥️ <a href="https://nuextract.ai/">API / Platform</a>&nbsp&nbsp | &nbsp&nbsp📑 <a href="https://numind.ai/blog">Blog</a>&nbsp&nbsp | &nbsp&nbsp🗣️ <a href="https://discord.gg/3tsEtJNCDe">Discord</a> &nbsp&nbsp | &nbsp&nbsp🛠️  <a href="https://github.com/numindai/nuextract">Github</a>
+</p>
+<section>
+  <h3> This space is a demo for <a href="https://huggingface.co/numind/NuExtract-2.0-4B" target="_blank">NuExtract-2.0-4B</a> </h3>
+  <h3> You can also check: <a href="https://huggingface.co/numind/NuExtract-2.0-2B" target="_blank">NuExtract-2.0-2B</a> and <a href="https://huggingface.co/numind/NuExtract-2.0-8B" target="_blank">NuExtract-2.0-8B</a> and our top performing model via the <a href="https://nuextract.ai/">API / Platform</a> </h3>
+    <h1>NuExtreact-2.0</h1>
+  <p>NuExtract 2.0 is a family of models trained specifically for structured information extraction tasks. It supports both multimodal inputs and is multilingual.</p>
+<p> To use the model, provide an input text/image and a JSON template describing the information you need to extract. The template should be a JSON object, specifying field names and their expected type. </p>
+  <article>
+    <h3>Supported Template Types</h3>
+    <ul>
+      <li><code>verbatim-string</code> — extract text exactly as it appears.</li>
+      <li><code>string</code> — generic text, with possible paraphrasing.</li>
+      <li><code>integer</code> — whole number.</li>
+      <li><code>number</code> — decimal or whole number.</li>
+      <li><code>date-time</code> — ISO 8601 date format.</li>
+      <li><code>boolean</code> — True or False.</li>
+      <li>Array of any type above (e.g. <code>["string"]</code>).</li>
+      <li><code>enum</code> — one value from a predefined list (e.g. <code>["yes", "no", "maybe"]</code>).</li>
+      <li><code>multi-label</code> — multiple values from a list (e.g. <code>[["A", "B", "C"]]</code>).</li>
+    </ul>
+    <p> You can specify any nested strucure, such as object inside object or list of object </p>
+    <p>If no relevant information is found, the model returns <code>null</code> or <code>[]</code>.</p>
+  </article>
+  <article>
+    <h3>Example Template</h3>
+    <pre><code>{
+  "first_name": "verbatim-string",
+  "last_name": "verbatim-string",
+  "description": "string",
+  "age": "integer",
+  "classes": [
+      {
+          "name" : "verbatim-string",
+          "professors" : ["verbatim-string"],
+          "gpa": number",
+      }
+  ],
+  "average_gpa": "number",
+  "birth_date": "date-time",
+  "nationality": ["France", "England", "Japan", "USA", "China"],
+  "languages_spoken": [["English", "French", "Japanese", "Mandarin", "Spanish"]]
+}</code></pre>
+  </article>
+  <strong>You can also provide a description of what you want to extract, use a non-JSON format (e.g. YAML, Pydantic) or even an example of input text. The model will automatically update the template field and generate a compatible JSON template based on our typing system.</strong>
+</section>
+<br>
+<section>
+  <ul><h4><strong>Model used in this demo:</strong> <a href="https://huggingface.co/numind/NuExtract-2.0-4B" target="_blank">NuExtract-2.0-4B</a></h4></ul>
+  <i>⚠️ This demo restricts inputs to 10,000 tokens</i>
+</section>
+</body>
+</html>
+""")
+    templ = gr.Textbox(
+        lines=10,
+        label="Template (JSON or prompt)",
+        placeholder="JSON template or instruction/other format to generate a template",
+    )
+    with gr.Row(equal_height=False):
+        img = gr.Image(type="filepath", label="Input image (PNG)", scale=1)
+        txt = gr.Textbox(lines=10, label="Text", scale=1)
+    example_data = [
+        [
+            "examples/affiche.jpg",      # image file
+            "",                          # no text
+            """{
+    "movie_name": "verbatim-string",
+    "tagline": "verbatim-string",
+    "language": "string",
+    "motion_picture_association_rating": [
+        "G - General Audiences",
+        "PG - Parental Guidance Suggested",
+        "PG-13 – Parents Strongly Cautioned",
+        "R – Restricted",
+        "NC-17 – Adults Only",
+        "not provided"
+    ],
+    "movie_distribution_company": "verbatim-string",
+    "movie_production_company": ["verbatim-string"],
+    "theatre_release_date": "date-time",
+    "movie_director_name": "verbatim-string",
+    "actors_names": [
+        "verbatim-string"
+    ],
+    "award" : "verbatim-string",
+    "reviews": [
+        {
+            "critic_name": "verbatim-string",
+            "review_comment": "verbatim-string"
+        }
+    ],
+    "technologies": [
+        [
+            "Dolby Stereo",
+            "Dolby Digital",
+            "Dolby Stereo Digital",
+            "Dolby Atmos",
+            "Dolby Vision",
+            "Dolby Cinema",
+            "DTS",
+            "SDDS",
+            "IMAX",
+            "4DX"
+        ]
+    ]
+}"""
+        ],
+        [
+            None,                          # no image
+            """Provectus is a Silicon Valley-based Artificial Intelligence consultancy and solutions provider.
+At Provectus, we are obsessed with leveraging cloud, data, and AI to reimagine the way businesses operate, compete, and deliver customer value. With the wide range of AI solutions for various use cases and industry verticals, Provectus is recognized by technology analysts and top cloud vendors as a leading AI consultancy and solutions provider. We are transformational leaders for our clients and employees.
+Currently, we are looking for a highly motivated and self-driven Front End Developer.
+Join us!
+Briefly about your first project in our company:
+The customer is an online distributor of menswear and a leader in the US market that works with Dolce & Gabbana, Calvin Klein, Ralph Lauren. We are developing a system for automating the processes of renting, buying, selling, and ordering suits online.
+Team:
+Frontend, Backend, and Mobile Engineers.
+Team Lead and PM are on the customer side in the United States.
+Requirements
+Software Engineering or related field with 3+ years of professional software development experience in Frontend technology such as ReactJs, AngularJs
+Deep understanding of front end development fundamentals including JavaScript, CSS, HTML and strong skills in React.js and its core principles, React.js workflows (such as Flux or Redux), Typescript, webpack, Babel, npm.
+Strong skills working with REST-based APIs and JSON data structures
+Hands-on experience working  experience with source control system Git
+Experience working with micro-frontend using web components
+Strong analytical and debugging skills
+At least an Intermediate level of English.
+Responsibilities
+Implementing best practices and technical solutions in process of migration from Angular to React
+Working on new functionality
+Taking ownership of business requirements and design, implement, test solutions
+Write a professional, performant, high-quality code that will support a scaling business
+Ask the right questions and think deeply about building solutions that support both - short-term and long-term goals
+Proactively participate in the agile development process sprints, providing effort estimates, commitments, and feedback to tasks.""",   # text
+            """{
+    "company": "verbatim-string",
+    "industry": "string",
+    "position": "string",
+    "contract_type": "string",
+    "location": "string",
+    "remote": [
+        "yes",
+        "no",
+        "hybrid"
+    ],
+    "education": "string",
+    "years_of_experience": "string",
+    "required_skills": [
+        "string"
+    ],
+    "responsibilities": [
+        "string"
+    ],
+    "salary": "string",
+    "benefits": [
+        "string"
+    ],
+    "language_skills": [
+        {
+            "language": "verbatim-string",
+            "level": "string"
+        }
+    ]
+}"""
+        ],
+    ]
+    warn = gr.Markdown(visible=False)
+    run_btn = gr.Button("Run", variant="primary")
+    out_json = gr.Textbox(
+        lines=14,
+        label="Extraction Output (JSON)",
+        show_copy_button=True,
+    )
+    run_btn.click(
+        fn=run_model,
+        inputs=[img, txt, templ],
+        outputs=[out_json, templ, warn],
+    )
+    gr.Examples(
+        examples=example_data,
+        inputs=[img, txt, templ],
+        label="🔍 Click an example to pre-fill the inputs",
+        cache_examples=False,
+    )
+demo.launch(debug=True, share=True)