|
|
--- |
|
|
license: mit |
|
|
tags: |
|
|
- generated_from_trainer |
|
|
model-index: |
|
|
- name: FolkGPT |
|
|
results: [] |
|
|
datasets: |
|
|
- vicclab/fairy_tales |
|
|
language: |
|
|
- en |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
# FolkGPT |
|
|
|
|
|
This model is a fine-tuned version of [gpt2](https://huggingface.co/gpt2) on vicclab/fairy_tales dataset. |
|
|
|
|
|
## Model description |
|
|
|
|
|
This model is the result of fine-tuning gpt2 on a dataset of fairy tales from various cultures. |
|
|
|
|
|
## Intended uses & limitations |
|
|
|
|
|
The idea behind this is to generate text in the fashion of fairy tales written in the 18th and 19th centuries. |
|
|
|
|
|
Why? Fairy tales seemed an appropriate application for text generation, as stories are usually short(ish), |
|
|
self-contained, and easy to read. |
|
|
|
|
|
## Training and evaluation data |
|
|
|
|
|
Trained on the vicclab/fairy_tales dataset. The dataset consists of a number of texts which |
|
|
were downloaded from Project Gutenberg, and then edited to remove all text except for the |
|
|
stories themselves. These were then all concatenated into a text file and pushed to HF at |
|
|
https://huggingface.co/datasets/vicclab/fairy_tales. The latest update to the dataset, which |
|
|
was used in the training of this model, was created and uploaded on February 26th, 2023. |
|
|
Texts used [and token count after removing boilerplate text]: |
|
|
https://www.gutenberg.org/files/2591/2591-0.txt [102927 tokens] |
|
|
https://www.gutenberg.org/files/503/503-0.txt [138353 tokens] |
|
|
https://www.gutenberg.org/cache/epub/69739/pg69739.txt [51035 tokens] |
|
|
https://www.gutenberg.org/files/2435/2435-0.txt [98791 tokens] |
|
|
https://www.gutenberg.org/cache/epub/7871/pg7871.txt [49410 tokens] |
|
|
https://www.gutenberg.org/files/8933/8933-0.txt [178622 tokens] |
|
|
gutenberg.org/cache/epub/30834/pg30834.txt [58359 tokens] |
|
|
https://www.gutenberg.org/cache/epub/68589/pg68589.txt [39815 tokens] |
|
|
https://www.gutenberg.org/cache/epub/34453/pg34453.txt [69365 tokens] |
|
|
gutenberg.org/cache/epub/8653/pg8653.txt [35351] |
|
|
|
|
|
[Total tokens in actual dataset: 1002654 tokens] |
|
|
|
|
|
|
|
|
## Training procedure |
|
|
The dataset was loaded, sampling by paragraph. From here, the dataset was split into a training dataset |
|
|
and a validation dataset in an 80-20 split. These were then tokenized. The model was set up, and the trainer |
|
|
was instantiated with the training_arguments listed below. Then, the training took place. |
|
|
|
|
|
### Training hyperparameters |
|
|
|
|
|
The following hyperparameters were used during training: |
|
|
- learning_rate: 0.0005 |
|
|
- train_batch_size: 2 |
|
|
- eval_batch_size: 2 |
|
|
- seed: 42 |
|
|
- gradient_accumulation_steps: 2 |
|
|
- total_train_batch_size: 4 |
|
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
|
- lr_scheduler_type: cosine |
|
|
- lr_scheduler_warmup_steps: 1000 |
|
|
- num_epochs: 1 |
|
|
- mixed_precision_training: Native AMP |
|
|
|
|
|
### Training results |
|
|
|
|
|
|
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- Transformers 4.26.1 |
|
|
- Pytorch 1.13.1+cu116 |
|
|
- Datasets 2.10.0 |
|
|
- Tokenizers 0.13.2 |
