MohamedBelal-AI/deberta-v3-base-conll2003-ner

Named Entity Recognition — DeBERTa-v3-base on CoNLL-2003

Fine-tuned microsoft/deberta-v3-base for English NER (PER, ORG, LOC, MISC) on CoNLL-2003.

Results (entity-level, seqeval)

• Test F1: 92.48 (precision 91.58, recall 93.40)
• Validation F1 (official dev set, best epoch 5): 96.21
• Test F1 with fully lowercased input: 88.75
• PER: F1 97.58 (support 1617)
• ORG: F1 91.01 (support 1661)
• LOC: F1 93.90 (support 1668)
• MISC: F1 81.30 (support 702)

Training setup

• Official CoNLL-2003 train/validation/test splits, -DOCSTART- markers removed (14,041 / 3,250 / 3,453 sentences)
• First-sub-word label alignment; entity-level evaluation with seqeval
• AdamW, lr 2e-05, weight decay 0.01, linear warmup (10%) + decay, gradient clipping 1.0, seed 42
• Lowercase augmentation on 20% of training sentences; best checkpoint chosen by validation F1 with early stopping

Limitations

• English only; trained on 1990s Reuters newswire, so modern or domain-specific text (tech, social media) may degrade.
• Performance depends on capitalization; see the lowercased-input score above.
• MISC is a heterogeneous category and the weakest type; ORG/LOC ambiguity (e.g. city names used for teams) is the most common confusion.
• Inputs longer than 256 sub-word tokens are truncated.