greek-ner-numeric-greekbert

Numeric-entity member of the FinNLP 2026 Subtask 1 (GRFinNUM) submission's ensemble. Fine-tuned nlpaueb/bert-base-greek-uncased-v1 as a token classifier.

Labels: ΧΡΗΜΑΤΑ (money) / ΠΟΣΟΣΤΑ (percent) / ΧΡΟΝΙΚΑ (temporal) / ΠΟΣΟΤΗΤΕΣ (quantity) / ΑΛΛΑ (other numeric)

Usage

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

model_id = "paragdakle/greek-ner-numeric-greekbert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id).eval()

text = "Στις 08.11.2019, η ΟΠΑΠ INVESTMENT LTD ήρθε σε συμφωνία..."
enc = tokenizer(text, return_offsets_mapping=True, return_tensors="pt")
offsets = enc.pop("offset_mapping")[0]
with torch.no_grad():
    logits = model(**enc).logits
tags = [model.config.id2label[i] for i in logits.argmax(-1)[0].tolist()]
# pair `tags` with `offsets` to recover character-span entities (see the
# submission's ner_utils.encoder_predict for the full decode)

Training data

400 sentences: the public GRFinNUM train (320) + validation (80) splits combined, since the official hidden test set now serves the role validation played during model selection.

Results

• validation split, this member alone: micro-F1 0.959, macro-F1 0.903
• validation split, 3-member weighted-vote ensemble (this + CRF + regex): micro-F1 0.982

Limitations

• Part of a submitted ensemble, not intended as a standalone SOTA tagger -- see the submission README for how this member is combined with the others.
• Small labeled training pool for the numeric task (320-400 sentences); silver-standard, general-domain data for the WikiANN-based text model.
• Financial-report Greek only (Athens Stock Exchange annual reports); not validated on other registers or domains.