greek-ner-numeric-crf

greek-ner-numeric-crf

Numeric-entity member of the FinNLP 2026 Subtask 1 (GRFinNUM) submission's ensemble. A linear-chain CRF (sklearn-crfsuite) over handcrafted shape/context features -- not a HF-native architecture, hosted as a plain joblib-pickled file.

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

Not a transformers-loadable model — a sklearn-crfsuite CRF object, pickled with joblib. Download and load it directly:

import joblib
from huggingface_hub import hf_hub_download

path = hf_hub_download("paragdakle/greek-ner-numeric-crf", "numeric_crf_model.joblib")
crf = joblib.load(path)

# crf.predict() expects one list of per-token feature dicts per sentence --
# the exact feature function is in the submission's ner_utils.crf_features,
# since predictions are meaningless unless features match training exactly.

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 0.947, macro-F1 0.911
• validation split, 3-member weighted-vote ensemble (this + GreekBERT + regex): micro-F1 0.982

Limitations

Part of a submitted ensemble, not intended as a standalone tagger — see the submission README for how this member is combined with the others.