Multilingual SpanMarker NER: mDeBERTa-v3 on WikiANN (7 Languages)
Developed by: Rishabh Kumar
Model Name: Rishabh157/spanmarker-wikiann-mdeberta
Base Architecture: microsoft/mdeberta-v3-base
This repository provides an enterprise-grade, state-of-the-art multilingual Named Entity Recognition (NER) model developed by Rishabh Kumar, combining microsoft/mdeberta-v3-base with the SpanMarker (PL-Marker) span-classification framework, fine-tuned across 7 major languages from WikiANN for 3 full epochs (26,934 steps).
Model Overview
- Developer / Creator: Rishabh Kumar (@Rishabh157)
- Base Architecture:
microsoft/mdeberta-v3-base(12 layers, 768 hidden dimension, disentangled relative positional attention). - Task Formulation: SpanMarker (Candidate phrase classification with boundary marker tokens [S_i, E_j] and Greedy Non-Maximum Suppression).
- Tuning Strategy: Full-Parameter Fine-Tuning (all 278.8M parameters active).
- Supported Languages: German (
de), English (en), Spanish (es), French (fr), Italian (it), Portuguese (pt), Swedish (sv). - Entity Taxonomy:
PER(Person)ORG(Organization)LOC(Location / Geopolitical Entity)
Empirical Benchmark Performance
Evaluated on the full official multilingual WikiANN validation set (70,000 sentences across all 7 languages, 10,000 sentences per language; 93,466 total gold entities) using strict exact-span seqeval matching:
Overall 3-Epoch Validation Results
| Metric | Score | Notes |
|---|---|---|
| Overall F1 Score | 76.75% | +15.13 point gain over Epoch 1 (61.62%) |
| Extraction Precision | 88.85% | Consistently high precision anchor, zero false-alarm hallucinations |
| Extraction Recall | 67.55% | Surged from 48.49% to 67.55% via multi-epoch statistical exposure |
| Sequence Accuracy | 87.37% | Exact token and span boundary classification |
| Validation Loss | 0.0421 | Steady monotonic decrease across all 3 epochs |
Per-Class Performance Breakdown (70,000 Sentences, 93,466 Entities)
| Entity Type | Semantic Category | Precision | Recall | F1 Score | Support (# Entities) |
|---|---|---|---|---|---|
PER |
Person / Names | 92.06% | 73.07% | 81.47% | 31,416 |
LOC |
Location / Geopolitical | 88.73% | 65.09% | 75.09% | 33,627 |
ORG |
Organization | 85.26% | 64.36% | 73.35% | 28,423 |
| Overall | Micro-Average | 88.85% | 67.55% | 76.75% | 93,466 |
Note on "Support (# Entities)": In Named Entity Recognition benchmarks, Support indicates the exact number of ground-truth entity spans belonging to that category in the validation dataset (70,000 sentences). It defines the sample size over which Recall and F1 are computed, providing transparency into entity frequency across the corpus.
Epoch-by-Epoch Convergence Trajectory
| Epoch | Step | Eval Loss | Precision | Recall | F1 Score | Accuracy |
|---|---|---|---|---|---|---|
| 1.0 | 8,978 | 0.0677 | 84.50% | 48.49% | 61.62% | 80.38% |
| 2.0 | 17,956 | 0.0490 | 87.77% | 61.86% | 72.57% | 84.64% |
| 3.0 | 26,934 | 0.0421 | 88.85% | 67.55% | 76.75% | 87.37% |
Quickstart & Inference
from span_marker import SpanMarkerModel
# Load model directly from Hugging Face Hub
model = SpanMarkerModel.from_pretrained("Rishabh157/spanmarker-wikiann-mdeberta")
# Multilingual inference examples
sentences = [
"Katia and Maurice Krafft died at Mount Unzen in Japan in 1991.",
"An der Yale University sind der Benny Goodman Nachlass und in Harvard der von Eubie Blake.",
"Barra do Garças Futebol Clube é um clube brasileiro de futebol da cidade de Barra do Garças."
]
predictions = model.predict(sentences)
for sentence, preds in zip(sentences, predictions):
print(f"\nSentence: {sentence}")
for entity in preds:
print(f" - [{entity['label']}] '{entity['span']}' (char {entity['char_start_index']}:{entity['char_end_index']}, score: {entity['score']:.3f})")
Training Configuration
- Batch Size: 16 per device (effective batch size 16)
- Steps: 26,934 steps (3 full epochs, ~4.17 hours runtime)
- Precision: FP16 mixed precision
- Sequence Length: 256 tokens max length, 16 words max entity span
- Optimizer: AdamW (lr=3.5e-5, cosine schedule, linear warmup)
Author & Citation
This model was trained, evaluated, and published by Rishabh Kumar (@Rishabh157).
If you use this model or refer to this work, please cite:
@misc{rishabhkumar2026_spanmarker_wikiann,
author = {Rishabh Kumar},
title = {SpanMarker Multilingual NER with mDeBERTa-v3 on WikiANN},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/Rishabh157/spanmarker-wikiann-mdeberta}}
}