metadata
license: apache-2.0
datasets:
- conll2003
- ai4privacy/pii-masking-200k
language:
- en
metrics:
- accuracy
- f1
library_name: transformers
pipeline_tag: token-classification
Model Details
Model Description
This model is electra-small finetuned for NER prediction task. The model currently predicts three entities which are given below.
- Location
- Person
- Organization
- Developed by:
விபின் (Vipin) - Model type: Google's electra small discriminator
- Language(s) (NLP): English
- License: Apache 2.0
- Finetuned from model [optional]: Google's electra small discriminator
Model Sources [optional]
Uses
This model uses tokenizer that is from distilbert family. So the model may predict wrong entities for same word (different sub word). Use 'aggregation_strategy' to "max" when using transformer's pipeline. for example 'ashwin ::" ash" => Person win => Location
Out-of-Scope Use
May not work well for some long sentences.
How to Get Started with the Model
Use the code below to get started with the model.
from transformers import AutoModelForTokenClassification, AutoTokenizer
from transformers import pipeline
model = AutoModelForTokenClassification.from_pretrained("rv2307/electra-small-ner")
tokenizer = AutoTokenizer.from_pretrained("rv2307/electra-small-ner")
nlp = pipeline("ner",
model=model,
tokenizer=tokenizer,device="cpu",
aggregation_strategy = "max")
Training Details
Training Procedure
This model is trained for 6 epoch in 3e-4 lr.
[39168/39168 41:18, Epoch 6/6]
Step Training Loss Validation Loss Precision Recall F1 Accuracy
10000 0.086300 0.088625 0.863476 0.876271 0.869827 0.972581
20000 0.059800 0.079611 0.894612 0.884521 0.889538 0.976563
30000 0.050400 0.074552 0.895812 0.902591 0.899188 0.978380
Evaluation
Validation loss is 0.07 for this model