Ihor/GLiNER-FUNSD

Dataset derived from FUNSD for training GLiNER-based multi-modal models.

Processing script:

from datasets import load_dataset
import shutil
import json
import os

dataset = load_dataset("nielsr/funsd")
raw_labels = dataset['train'].features['ner_tags'].feature.names

gliner_pdf_dataset = []
output_dir = "gliner_funsd"
os.makedirs(output_dir, exist_ok=True)
os.makedirs(os.path.join(output_dir, 'images'), exist_ok=True)

def process_dataset(example):
    tokens = example['words']
    bboxes  = example['bboxes']
    img_src = example['image_path']

    ner_spans = []
    prev_label = 'O'
    prev_idx   = None

    for idx, tag_id in enumerate(example['ner_tags']):
        lbl = raw_labels[tag_id]
        if lbl == 'O':
            if prev_label != 'O':
                ner_spans.append([prev_idx, idx - 1, prev_label])
                prev_label = 'O'
            continue

        _, label = lbl.split('-', 1)
        if prev_label != label:
            if prev_label != 'O':
                ner_spans.append([prev_idx, idx - 1, prev_label])
            prev_idx   = idx
            prev_label = label

    if prev_label != 'O':
        ner_spans.append([prev_idx, len(example['ner_tags']) - 1, prev_label])

    fname      = os.path.basename(img_src)
    img_dst    = os.path.join(output_dir, 'images', fname)
    shutil.copy(img_src, img_dst)

    gliner_pdf_dataset.append({
        "tokenized_text":     tokens,
        "bboxes":     bboxes,
        "ner":  ner_spans,
        "image_path": img_dst
    })

for ex in dataset['train']:
    process_dataset(ex)

with open(os.path.join(output_dir, 'data.json'), 'w') as f:
    json.dump(gliner_pdf_dataset, f)