This is a roberta pre-trained version on the CodeSearchNet dataset for Java Mask Language Model mission.

To load the model: (necessary packages: !pip install transformers sentencepiece)

from transformers import AutoTokenizer, AutoModelWithLMHead, pipeline
tokenizer = AutoTokenizer.from_pretrained("dbernsohn/roberta-java")
model = AutoModelWithLMHead.from_pretrained("dbernsohn/roberta-java")

fill_mask = pipeline(

You can then use this model to fill masked words in a Java code.

code = """
String[] cars = {"Volvo", "BMW", "Ford", "Mazda"};
for (String i : cars) {

pred = {x["token_str"].replace("Ġ", ""): x["score"] for x in fill_mask(code)}
sorted(pred.items(), key=lambda kv: kv[1], reverse=True)
# [('println', 0.32571351528167725),
# ('get', 0.2897663116455078),
# ('remove', 0.0637081190943718),
# ('exit', 0.058875661343336105),
# ('print', 0.034190207719802856)]

The whole training process and hyperparameters are in my GitHub repo

Created by Dor Bernsohn

Downloads last month
Hosted inference API

Mask token: <mask>