distilbert-base-uncased-finetuned-ner-exp_B

<!-- This model card has been generated automatically according to the information Keras had access to. You should probably proofread and complete it, then remove this comment. -->

i-be-snek/distilbert-base-uncased-finetuned-ner-exp_B

This model is a fine-tuned version of distilbert-base-uncased on the English subset of the NER Babelscape/multinerd dataset.

It achieves the following results on the validation set:

• Train Loss: 0.0084
• Validation Loss: 0.0587
• Train Precision: 0.9185
• Train Recall: 0.9240
• Train F1: 0.9213
• Train Accuracy: 0.9857
• Epoch: 2

Model description

distilbert-base-uncased-finetuned-ner-exp_B is a Named Entity Recognition model finetuned on distilbert-base-uncased. This model is uncased, so it makes no distinction between "sarah" and "Sarah". The dataset it was fine-tuned on was modified. Only five entities were considered: Person (PER), Animal (ANIM), Organization (ORG), Location (LOC), and Disease (DIS). The dataset was modified further so that all other named entities not included in this list were swapped with the '0' label ID. Tokens IDs were also re-indexed.

Training and evaluation data

This model has been evaluated on the English subset of the test set of Babelscape/multinerd with modifications where all tags other than (0) Person (PER), Animal (ANIM), Organization (ORG), Location (LOC), and Disease (DIS) were replaced with the 'O' tag. The label indices were also reset and the dataset was transformed accordingly. You can preprocess the dataset in the same way with any custom set of tags using the script in this GitHub repository

Evaluation results

Training procedure

All scripts for training can be found in this GitHub repository. The model had early stopped watching its val_loss.

Training hyperparameters

The following hyperparameters were used during training:

• optimizer:

•   {
      "name": "AdamWeightDecay",
      "learning_rate": 2e-05,
      "decay": 0.0,
      "beta_1": 0.9,
      "beta_2": 0.999,
      "epsilon": 1e-07,
      "amsgrad": False,
      "weight_decay_rate": 0.0,
  }
  

• training_precision: float32

Training results

Epoch 0

Epoch 1

Epoch 2

Framework versions

• Transformers 4.35.2
• TensorFlow 2.14.0
• Datasets 2.15.0
• Tokenizers 0.15.0