app.py

import random
from gliner import GLiNER
import gradio as gr
from datasets import load_dataset

# Load the BL dataset as a streaming iterator
dataset_iter = load_dataset(
    "TheBritishLibrary/blbooks",
    split="train",
    streaming=True,  # Enable streaming
    trust_remote_code=True
).shuffle(seed=42)  # Shuffle added

# Load the model
model = GLiNER.from_pretrained("urchade/gliner_multi-v2.1", trust_remote_code=True)

def ner(text: str, labels: str, threshold: float, nested_ner: bool):
    # Convert user-provided labels (comma-separated string) into a list
    labels_list = [label.strip() for label in labels.split(",")]
    
    # Truncate the text to avoid length exceeding model limits (e.g., 384 tokens)
    max_length = 384
    truncated_text = text[:max_length]
    
    # Predict entities using the GLiNER model
    entities = model.predict_entities(truncated_text, labels_list, flat_ner=not nested_ner, threshold=threshold)
    
    # Prepare entities for color-coded display using gr.HighlightedText
    highlights = [{"start": ent["start"], "end": ent["end"], "entity": ent["label"]} for ent in entities]
    
    # Return both the highlighted text and the raw entities in JSON format
    return {
        "text": truncated_text,
        "entities": highlights
    }, entities  # Return both outputs: the first for HighlightedText, the second for JSON

with gr.Blocks(title="General NER with Color-Coded Output") as demo:
    gr.Markdown(
        """
        # GLiNER British Library Books Demo
        This demo selects a random text snippet from the British Library's books dataset and identifies entities using GLiNER (urchade/gliner_multi-v2.1). 
        """
    )
    
    # Display a random example
    input_text = gr.Textbox(
        value="Click on 'Get New Snippet' to load a piece of text from the British Library dataset",
        label="Text input",
        placeholder="Enter your text here",
        lines=5
    )
    refresh_btn = gr.Button("Get New Snippet")

    with gr.Row() as row:
        labels = gr.Textbox(
            value="Person, Location",  # Default example labels
            label="Labels",
            placeholder="Enter your labels here (comma separated)",
            scale=2,
        )
        threshold = gr.Slider(
            0,
            1,
            value=0.5,  # Adjusted to match the threshold used in the function
            step=0.01,
            label="Threshold",
            info="Lower the threshold to increase how many entities get predicted.",
            scale=1,
        )
        nested_ner = gr.Checkbox(
            value=False,
            label="Nested NER",
            info="Enable Nested NER?",
        )
    submit_btn = gr.Button("Find Entities!")

    # Define output components using HighlightedText for color-coded display
    output_highlighted = gr.HighlightedText(label="Predicted Entities")
    output_entities = gr.JSON(label="Entities")
    

    def get_new_snippet():
        # Preload several samples into a list
        max_length = 384  # Maximum length for snippets
        samples = [
            sample['text'][:max_length] for sample, _ in zip(dataset_iter, range(100))  # Truncate to max_length
        ]
        
        # Return a random snippet from the preloaded samples
        if samples:
            return random.choice(samples)
        
        return "No more snippets available."  # Return this if no valid snippets are found 
    
    # Connect refresh button
    refresh_btn.click(fn=get_new_snippet, outputs=input_text)
    
    # Connect submit button
    submit_btn.click(
        fn=ner,
        inputs=[input_text, labels, threshold, nested_ner],
        outputs=[output_highlighted, output_entities]
    )

demo.queue()
demo.launch(debug=True)