vesteinn's picture
Librarian Bot: Add base_model information to model (#3)
8e3ec1f
---
language:
- en
- is
- multilingual
license: agpl-3.0
tags:
- generated_from_trainer
datasets:
- conll2003
metrics:
- precision
- recall
- f1
- accuracy
base_model: vesteinn/XLMR-ENIS
model-index:
- name: XLMR-ENIS-finetuned-ner
results:
- task:
type: token-classification
name: Token Classification
dataset:
name: conll2003
type: conll2003
args: conll2003
metrics:
- type: precision
value: 0.9398313331170938
name: Precision
- type: recall
value: 0.9517943664285128
name: Recall
- type: f1
value: 0.9457750214207026
name: F1
- type: accuracy
value: 0.9853686150987764
name: Accuracy
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# XLMR-ENIS-finetuned-ner
This model is a fine-tuned version of [vesteinn/XLMR-ENIS](https://huggingface.co/vesteinn/XLMR-ENIS) on the conll2003 dataset.
It achieves the following results on the evaluation set:
- Loss: 0.0671
- Precision: 0.9398
- Recall: 0.9518
- F1: 0.9458
- Accuracy: 0.9854
## Model description
More information needed
## Intended uses & limitations
More information needed
## Training and evaluation data
More information needed
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 3
### Training results
| Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1 | Accuracy |
|:-------------:|:-----:|:----:|:---------------:|:---------:|:------:|:------:|:--------:|
| 0.2825 | 1.0 | 878 | 0.0712 | 0.9220 | 0.9379 | 0.9299 | 0.9815 |
| 0.0688 | 2.0 | 1756 | 0.0689 | 0.9354 | 0.9477 | 0.9415 | 0.9839 |
| 0.039 | 3.0 | 2634 | 0.0671 | 0.9398 | 0.9518 | 0.9458 | 0.9854 |
### Framework versions
- Transformers 4.10.3
- Pytorch 1.9.0+cu102
- Datasets 1.12.1
- Tokenizers 0.10.3