|
--- |
|
license: apache-2.0 |
|
base_model: bert-base-uncased |
|
tags: |
|
- generated_from_keras_callback |
|
model-index: |
|
- name: EconoBert |
|
results: [] |
|
datasets: |
|
- samchain/BIS_Speeches_97_23 |
|
language: |
|
- en |
|
pipeline_tag: fill-mask |
|
--- |
|
|
|
# EconoBert |
|
|
|
This model is a fine-tuned version of [bert-base-uncased](https://huggingface.co./bert-base-uncased) on this dataset: (https://huggingface.co./datasets/samchain/BIS_Speeches_97_23) |
|
It achieves the following results on the test set: |
|
|
|
- Accuracy for MLM task: 73% |
|
- Accuracy for NSP task: 95% |
|
|
|
## Model description |
|
|
|
The model is a simple fine-tuning of a base bert on a dataset specific to the domain of economics. It follows the same architecture and no resize_token_embeddings were required. |
|
|
|
## Intended uses & limitations |
|
|
|
This model should be used as a backbone for NLP tasks applied to the domain of economics, politics and finance. |
|
|
|
## Training and evaluation data |
|
|
|
The dataset used as a fine-tuning domain is : https://huggingface.co./datasets/samchain/BIS_Speeches_97_23 |
|
|
|
The dataset is made of 773k pairs of sentences, an half being negative pairs (meaning sequence A and B are not related) and the other half positive (sequence B follows sequence A). |
|
|
|
The test set is made of 136k pairs. |
|
|
|
## Training procedure |
|
|
|
The model has been fine tuned on 2 epochs, with a batch size of 64 and a sequence length of 128. I used Adam learning-rate with a value of 1e-5, |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- optimizer: {'name': 'Adam', 'weight_decay': None, 'clipnorm': None, 'global_clipnorm': None, 'clipvalue': None, 'use_ema': False, 'ema_momentum': 0.99, 'ema_overwrite_frequency': None, 'jit_compile': True, 'is_legacy_optimizer': False, 'learning_rate': 1e-05, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-07, 'amsgrad': False} |
|
- training_precision: float32 |
|
|
|
### Training results |
|
|
|
Training loss is 1.6046 on train set and 1.47 on test set. |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.31.0 |
|
- TensorFlow 2.12.0 |
|
- Datasets 2.13.1 |
|
- Tokenizers 0.13.3 |
