metadata

language:
  - en
license: mit
base_model: prajjwal1/bert-tiny
tags:
  - pytorch
  - BertForTokenClassification
  - named-entity-recognition
  - roberta-base
  - generated_from_trainer
metrics:
  - recall
  - precision
  - f1
  - accuracy
model-index:
  - name: bert-tiny-ontonotes
    results:
      - task:
          type: token-classification
        metrics:
          - type: accuracy
            value: 0.9476
            name: accuracy
          - type: precision
            value: 0.6817
            name: precision
          - type: accuracy
            value: 0.7193
            name: recall
          - type: accuracy
            value: 0.7
            name: F1
datasets:
  - tner/ontonotes5
library_name: transformers
pipeline_tag: token-classification

bert-tiny-ontonotes

This model is a fine-tuned version of prajjwal1/bert-tiny on the tner/ontonotes5 dataset. It achieves the following results on the evaluation set:

Loss: 0.1917
Recall: 0.7193
Precision: 0.6817
F1: 0.7000
Accuracy: 0.9476

How to use the Model

Using pipeline

from transformers import pipeline
import torch

# Initiate the pipeline
device = 0 if torch.cuda.is_available() else 'cpu'
ner = pipeline("token-classification", "arnabdhar/bert-tiny-ontonotes", device=device)

# use the pipeline
input_text = "My name is Clara and I live in Berkeley, California."
results = ner(input_text)

Intended uses & limitations

This model is fine-tuned for Named Entity Recognition task and you can use the model as it is or can use this model as a base model for further fine tuning on your custom dataset.

The following entities were fine-tuned on: CARDINAL, DATE, PERSON, NORP, GPE, LAW, PERCENT, ORDINAL, MONEY, WORK_OF_ART, FAC, TIME, QUANTITY, PRODUCT, LANGUAGE, ORG, LOC, EVENT

Training and evaluation data

The dataset has 3 partitions, train, validation and test, all the 3 partitions were combined and then a 80:20 train-test split was made for finet uning process. The following ID2LABEL mapping was used.

{
    "0": "O",
    "1": "B-CARDINAL",
    "2": "B-DATE",
    "3": "I-DATE",
    "4": "B-PERSON",
    "5": "I-PERSON",
    "6": "B-NORP",
    "7": "B-GPE",
    "8": "I-GPE",
    "9": "B-LAW",
    "10": "I-LAW",
    "11": "B-ORG",
    "12": "I-ORG",
    "13": "B-PERCENT",
    "14": "I-PERCENT",
    "15": "B-ORDINAL",
    "16": "B-MONEY",
    "17": "I-MONEY",
    "18": "B-WORK_OF_ART",
    "19": "I-WORK_OF_ART",
    "20": "B-FAC",
    "21": "B-TIME",
    "22": "I-CARDINAL",
    "23": "B-LOC",
    "24": "B-QUANTITY",
    "25": "I-QUANTITY",
    "26": "I-NORP",
    "27": "I-LOC",
    "28": "B-PRODUCT",
    "29": "I-TIME",
    "30": "B-EVENT",
    "31": "I-EVENT",
    "32": "I-FAC",
    "33": "B-LANGUAGE",
    "34": "I-PRODUCT",
    "35": "I-ORDINAL",
    "36": "I-LANGUAGE"
  }

Training procedure

The model was finetuned on Google Colab with a NVIDIA T4 GPU with 15GB of VRAM. It took around 5 minutes to fine tune and evaluate the model with 6000 steps of total training steps. For more details, you can look into the Weights & Biases log history.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 8e-05
train_batch_size: 32
eval_batch_size: 160
seed: 75241309
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
training_steps: 6000

Training results

Training Loss	Epoch	Step	Validation Loss	Recall	Precision	F1	Accuracy
0.4283	0.31	600	0.3864	0.4561	0.4260	0.4405	0.9058
0.3214	0.63	1200	0.2865	0.5865	0.5485	0.5669	0.9265
0.2886	0.94	1800	0.2439	0.6432	0.6165	0.6295	0.9354
0.2511	1.25	2400	0.2233	0.6765	0.6250	0.6497	0.9389
0.2224	1.56	3000	0.2088	0.6878	0.6642	0.6758	0.9433
0.2181	1.88	3600	0.2001	0.7105	0.6684	0.6888	0.9451
0.215	2.19	4200	0.1954	0.7140	0.6795	0.6963	0.9469
0.1907	2.5	4800	0.1934	0.7169	0.6776	0.6967	0.9470
0.209	2.82	5400	0.1918	0.7185	0.6812	0.6994	0.9475
0.2073	3.13	6000	0.1917	0.7193	0.6817	0.7000	0.9476

Framework versions

Transformers 4.36.2
Pytorch 2.1.0+cu121
Datasets 2.15.0
Tokenizers 0.15.0