|
--- |
|
license: mit |
|
tags: |
|
- generated_from_trainer |
|
model-index: |
|
- name: gpt2-finetuned-scientific-articles |
|
results: [] |
|
--- |
|
|
|
This repository is the submission for the final project for BF510 [Institutional Racism in Health and Science](http://irhs.bu.edu/) for Shariq Madha. |
|
|
|
To see Jupyter detailing how this model was produced, as well as the motivation behind it, go [here](https://github.com/ssmadha/BF510-final-project/). |
|
|
|
To try this out yourself, enter a prompt in the textbox to the right and hit compute (it may take a minute for the first to process, but subsequent results should be quick). |
|
|
|
# gpt2-finetuned-scientific-articles |
|
|
|
This model is a fine-tuned version of [gpt2](https://huggingface.co/gpt2) on scientific articles about algorithmic bias. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 2.3793 |
|
|
|
## Model description |
|
|
|
This model is a casual language modeling GPT2 fine-tuned on scientific articles about algorithmic bias, in an attempt to showcase an example about correcting for algorithmic bias. |
|
|
|
## Intended uses & limitations |
|
|
|
This model is intended for prompts about algorithms and bias. Other prompts will yield results, but they are less likely to be influenced by the fine-tuning. |
|
|
|
## Training and evaluation data |
|
|
|
This model is trained on fully freely accessible articles obtained from a PubMed Central search on algorithmic bias. The pmc_result_algorithmicbias.txt file contains the list of PMC's used. Due to technical and time limitations, only fine-tuned on the introduction sections, but training on other sections is planned. |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 2e-05 |
|
- train_batch_size: 4 |
|
- eval_batch_size: 4 |
|
- seed: 42 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: linear |
|
- num_epochs: 2 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | |
|
|:-------------:|:-----:|:----:|:---------------:| |
|
| 2.5293 | 1.0 | 1071 | 2.3892 | |
|
| 2.4821 | 2.0 | 2142 | 2.3793 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.14.0.dev0 |
|
- Pytorch 1.10.0+cu102 |
|
- Datasets 1.15.1 |
|
- Tokenizers 0.10.3 |
|
|
