BERT fine-tuned for Named Entity Recognition (CoNLL-2003)
A fine-tuned version of bert-base-cased
for Named Entity Recognition (NER), trained on the CoNLL-2003 English dataset as part
of working through the Hugging Face LLM Course,
Chapter 7.
Model details
| Attribute | Value |
|---|---|
| Base model | bert-base-cased |
| Architecture | Transformer Encoder (BERT) |
| Task | Token Classification (NER) |
| Training dataset | CoNLL-2003 (English) |
| Training epochs | 3 |
| Learning rate | 2e-5 |
| Weight decay | 0.01 |
| Hardware | Google Colab (T4 GPU) |
Entity types
The model recognises four entity types in IOB2 format:
| Label | Description |
|---|---|
| PER | Person |
| ORG | Organisation |
| LOC | Location |
| MISC | Miscellaneous |
Usage
from transformers import pipeline
ner = pipeline(
"token-classification",
model="AlexStamp/bert-finetuned-ner",
aggregation_strategy="simple"
)
ner("Alexis works at CERN in Switzerland.")
Training procedure
Fine-tuning was performed using the Hugging Face Trainer API with
DataCollatorForTokenClassification and evaluated using the seqeval
library, which computes entity-level F1 — stricter than token-level accuracy
since the entire entity span must be correctly identified.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss | Validation Precision | Validation Recall | Validation F1 | Validation Accuracy |
|---|---|---|---|---|---|---|---|
| 0.0759 | 1.0 | 1756 | 0.0651 | 0.8905 | 0.9310 | 0.9103 | 0.9812 |
| 0.0355 | 2.0 | 3512 | 0.0681 | 0.9321 | 0.9473 | 0.9397 | 0.9853 |
| 0.0224 | 3.0 | 5268 | 0.0599 | 0.9319 | 0.9507 | 0.9412 | 0.9867 |
Framework versions
- Transformers 5.12.0
- Pytorch 2.11.0+cu128
- Datasets 4.0.0
- Tokenizers 0.22.2
Limitations
- Trained on English news wire text (Reuters corpus); may generalise poorly to other domains or languages
bert-base-casedis case-sensitive by design, which is appropriate for NER but means casing errors in input text can degrade performance
Notes
This model was trained as a portfolio exercise. The base model choice
(bert-base-cased over bert-base-uncased) is deliberate — NER is
case-sensitive since capitalisation is a strong signal for entity detection.