Elpapudex/Ner-full

u2t01-bert-ner

Training Data

• Dataset: lhoestq/conll2003
• Base Model: bert-base-uncased
• Hyperparameters:
  • head_lr: 1e-3
  • body_lr: 2e-5
  • seed: 42
• Tokenization & Labeling Strategy:
  • BIO scheme
  • First-subword labeling
  • -100 mask for continuation and special tokens

Metrics

{
  "loss": 0.0940466895699501,
  "token_accuracy": 0.9868920787643886,
  "precision": 0.925249169435216,
  "recall": 0.937868328001347,
  "f1": 0.9315160130445689,
  "runtime": 8.1007,
  "samples_per_second": 401.199,
  "steps_per_second": 3.21,
  "epoch": 3.0,
  "method": "full",
  "trainable_params": 108898569,
  "total_train_seconds": 309.33441281318665
}

Headline Metric: Entity-level F1 = 0.9315

Exaple Usage

from transformers import pipeline

# 1. Initialize the NER pipeline
ner = pipeline(
    task="token-classification",
    model="Elpapudex/Ner-full",
    aggregation_strategy="simple"  # Groups sub-tokens into complete entity names
)

# 2. Text to test
text = "Sundar Pichai announced that Google is opening a new AI hub in London next October."

# 3. Run inference
results = ner(text)

# 4. Display detected entities
for entity in results:
    print(f"Entity: {entity['word']} | Label: {entity['entity_group']} | Confidence: {entity['score']:.2%}")

Limitations

• English-only.
• Trained on the CoNLL-2003 newswire domain.
• BIO scheme dependent.
• May not generalize well to specialized domains such as legal, scientific, conversational, or historical text.

References

• Devlin et al., BERT, arXiv:1810.04805 §5.3