uzbek-ner-xlmr-large

uzbek-ner-xlmr-large — Uzbek NER (XLM-R-large fine-tuned on Uzbek-NER-Gold)

Token-classification model for Uzbek Named Entity Recognition: FacebookAI/xlm-roberta-large fine-tuned on 3,776 Gold-train sentences (uznlp-uz/uzbek_NER, seed-7 split). SUFFIX-INCLUSIVE convention — entity spans include Uzbek case/possessive/plural suffixes as annotated in Gold (e.g. Zarafshonga = B-LOC).

• License: MIT (base model FacebookAI/xlm-roberta-large is MIT).
• Training data license: Gold is CC-BY-4.0 (uznlp-uz/uzbek_NER, DOI 10.57967/hf/9399). No new human annotation in this release — weights + eval predictions only.
• This checkpoint = retrained seed 10 of a 5-seed run (closest-to-mean config; see paper §3). 5-seed means: 0.792 full-400 (range 0.777–0.802) · 0.809 LLM-200 (range 0.785–0.823). Single-seed numbers below are this checkpoint, verified locally on CPU after download.

Measured results (strict entity micro-F1, frozen test splits)

For comparison on the same splits: in-domain CRF 0.771/0.727 · GPT-5.6-Luna few-shot 0.621 · memorization 0.522 · GLiNER 0.495 · BERTbek-transfer 0.482. Full tables in the paper.

Tagset (17 labels)

O, B-PER I-PER, B-ORG I-ORG, B-LOC I-LOC, B-MISC I-MISC, B-TEMPORAL I-TEMPORAL, B-NUMERIC I-NUMERIC, B-WORK I-WORK, B-MONEY I-MONEY (BIO).

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

tok = AutoTokenizer.from_pretrained("UAzimov/uzbek-ner-xlmr-large")
model = AutoModelForTokenClassification.from_pretrained("UAzimov/uzbek-ner-xlmr-large")
model.eval()

words = ["Prezident", "Shavkat", "Mirziyoyev", "Samarqandga", "keldi", "."]
enc = tok(words, is_split_into_words=True, return_tensors="pt", truncation=True)
with torch.no_grad():
    pred = model(**enc).logits.argmax(-1)[0].tolist()
wids, tags, seen = enc.word_ids(), ["O"] * len(words), set()
id2l = model.config.id2label
for j, w in enumerate(wids):
    if w is None or w in seen or w >= len(words):
        continue
    seen.add(w)
    tags[w] = id2l[pred[j]]
print(list(zip(words, tags)))
# [('Prezident', 'O'), ('Shavkat', 'B-PER'), ('Mirziyoyev', 'I-PER'),
#  ('Samarqandga', 'B-LOC'), ('keldi', 'O'), ('.', 'O')]

First-subword label alignment (continuation subwords ignored at decode).

Training

• Base: FacebookAI/xlm-roberta-large · fresh token-classification head (17 labels)
• Data: 3,776 Gold-train sentences (seed-7 split of uznlp-uz/uzbek_NER)
• 4 epochs, lr 2e-5, per-device batch 8 × grad-accum 2 (effective 16), max_len 256, fp16, seed 10
• Fixed epochs, no dev set / no early stopping; evaluated once on frozen test-400
• Hardware: Kaggle T4 14.6 GB (~14 min train)

Limitations

• Single frozen split (seed 7); CIs quantify resampling noise, not split noise.
• 1/400 test sentences exactly duplicates a train sentence (dataset property).
• Pretraining contamination unknowable (XLM-R pretraining may include Gold-like Uzbek news).
• MONEY (n = 5) and other rare-type estimates are noisy by construction.

Citation

@misc{azimov2026uzbeknerxlmr,
  title = {uzbek-ner-xlmr-large: XLM-R-large fine-tuned for Uzbek NER on Uzbek-NER-Gold},
  author = {Azimov, Utkirbek},
  year = {2026},
  howpublished = {\url{https://huggingface.co/UAzimov/uzbek-ner-xlmr-large}},
  license = {MIT},
}

Base data: Elov, B. B. & Alaev, R. H. (2026). Uzbek NER Gold. https://huggingface.co/datasets/uznlp-uz/uzbek_NER (CC-BY-4.0, DOI 10.57967/hf/9399). Full benchmark paper with all 15 systems: see repo paper.md.