husnu's picture
Update README.md
4f9eccb
---
tags:
- generated_from_trainer
datasets:
- turkish squad v2
model-index:
- name: bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3TQUAD2-finetuned_lr-2e-05_epochs-3
results: []
---
<!-- 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. -->
# bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3TQUAD2-finetuned_lr-2e-05_epochs-3
This model is a fine-tuned version of [husnu/bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3](https://huggingface.co/husnu/bert-base-turkish-128k-cased-finetuned_lr-2e-05_epochs-3) on the turkish squad2 dataset.
It achieves the following results on the evaluation set:
- Loss: 1.9011
## 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: 8
- eval_batch_size: 8
- 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 |
|:-------------:|:-----:|:----:|:---------------:|
| 0.6404 | 1.0 | 2245 | 1.4524 |
| 0.403 | 2.0 | 4490 | 1.5638 |
| 0.2355 | 3.0 | 6735 | 1.9011 |
### Framework versions
- Transformers 4.15.0
- Pytorch 1.10.0+cu111
- Datasets 1.17.0
- Tokenizers 0.10.3