A Lossless Syntax Tree Generator with Zero-shot Error Correction
- We follow jam's pretraining procedure and use the same data to pretrain except we also use srcml to pretrain the models.
- In the finetuning stage, we finetune our models for 3 epochs.
- Our GitHub repo contains the code for reproduction using the same data.
Pretrained model parameters
| Hyperparameter | Description | Value |
|---|---|---|
| e | embedding dimensions | 1024 |
| L | number of layers | 24 |
| h | attention heads | 16 |
| c | block size / context length | 256 |
| b | batch size | 4 |
| a | accumulation steps | 32 |
| r | learning rate | 3e-5 |
| y | weight decay | 1e-5 |
| iter | iterations | 570000 |
Model files
| Filename | Description |
|---|---|
| ckpt.pt | A model file for finetuning |
| ckpt_base.pt | A model file for generating syntax tree with the error correction in zero-shot setting |
| ckpt_finetune.pt | A model finetuned with the syntatic error dataset |
- Note that you can adjust the batch size and accumulation steps based on your GPU memory. But, the batch size * accumulation steps should be 128.
- If you finetune your models with multiple GPUs, you can turn down accumulation steps. For example, if you finetune with 2 GPUs, you will need to half the accumulation steps.