|
--- |
|
language: |
|
- en |
|
tags: |
|
- transformers |
|
license: apache-2.0 |
|
datasets: |
|
- dvitel/hearthstone |
|
metrics: |
|
- exact_match |
|
- bleu |
|
- dvitel/codebleu |
|
- chrf |
|
--- |
|
|
|
Application of distilgpt2 to HearthStone card code synthesis. Dataset [dvitel/hearthstone](https://huggingface.co/datasets/dvitel/hearthstone) \ |
|
Article under consideration: [Abstract Syntax Networks for Code Generation and Semantic Parsing](https://aclanthology.org/P17-1105.pdf) \ |
|
We check if distilgpt2 could produce better results than ASNs from article. H0 model has minimal preprocessing of dataset where we normalize python code: |
|
|
|
```python |
|
def normalize(line:str): |
|
return line.strip().replace("§", "\n").replace(" ", "\t").replace("\\ ", "").replace("\n\n", "\n") |
|
``` |