|
# Tokenizer |
|
|
|
We trained our tokenizer using [sentencepiece](https://github.com/google/sentencepiece)'s unigram tokenizer. Then loaded the tokenizer as MT5TokenizerFast. |
|
|
|
## Model |
|
|
|
We used [MT5-base](https://huggingface.co/google/mt5-base) model. |
|
|
|
## Datasets |
|
|
|
We used [Code Search Net](https://huggingface.co/datasets/code_search_net)'s dataset and some scrapped data from internet to train the model. We maintained a list of datasets where each dataset had codes of same language. |
|
|
|
## Plots |
|
|
|
### Train loss |
|
|
|
![train loss](https://i.ibb.co/x53Wm8n/train-loss.png) |
|
|
|
### Evaluation loss |
|
|
|
![eval loss](https://i.ibb.co/McB2jnf/eval-loss.png) |
|
|
|
### Evaluation accuracy |
|
|
|
![eval accuracy](https://i.ibb.co/YDGhLdn/eval-accuracy.png) |
|
|
|
### Learning rate |
|
|
|
![learning rate](https://i.ibb.co/CMStzWv/learning-rate.png) |
|
|
|
## Fine tuning (WIP) |
|
|
|
We fine tuned the model with [CodeXGLUE code-to-code-trans dataset](https://huggingface.co/datasets/code_x_glue_cc_code_to_code_trans), and scrapper data. |
|
|