codecompletion model and input about codexglue

[celbree]

In the token-level completion scenario, when the inputs is "<s> package", assuming the ground truth is "org", but the model predicts "com". In the next step, the inputs should be "<s> package org". So it doesn't matter the model couldn't predict different tokens after "<s> package" because even though the model predict the wrong token, in the next step, the ground truth will be feed to model.

[celbree]

To the first question, CodeGPT-small-py-adaptedGPT2 is a pre-train model. And it need fine-tuned on PY150 or Github Java Corpus dataset for both token-level completion and line-level completion.
The second question. Yes. The inputs and ground truth is the same file. The ground truth is the "next token" given the context.

[Martine307]

Thank you. But I am still confused about the input for the token-level completion. When the input is "<s> package" and use beam search, the predicted next token should be same, right? How can i get the different tokens("org" or "com") in the following picture?