zoom_in Click to enlarge

InCoder

▲ ▼

language

description InCoder Overview

InCoder is a generative artificial intelligence language model developed by Meta, designed specifically for code generation and editing tasks. Unlike earlier models that only predicted the next text token, InCoder is uniquely trained to perform infilling, allowing it to insert code into the middle of existing files. It was trained on a massive dataset of publicly available source code and natural language to understand programming syntax and structure. The model is intended for software engineering researchers and developers exploring automated programming assistants.

help InCoder FAQ

Why can InCoder fill a hole in the middle of a function?

InCoder can generate code left-to-right and fill a masked span by using both the code before and after the gap. The 2022 paper tested this on missing lines, type prediction, docstrings, and variable renaming.

What model size is the main InCoder release?

The paper's primary model is InCoder-6.7B, a 6.7-billion-parameter Transformer. It also reports a smaller 1.3-billion-parameter model in its ablation experiments.

Which programming languages was InCoder tested on?

The paper's CodeXGLUE experiments cover Python, JavaScript, Ruby, Go, Java, and PHP. Its standard synthesis evaluation includes HumanEval, a Python function-generation benchmark.

Can InCoder edit an existing code file instead of writing from scratch?

Yes. You place mask sentinel tokens where code is missing, and the model generates replacement text for edits such as comments, type hints, and variable names.

swap_horiz

Looking for InCoder alternatives? Compare top competitors ranked & scored

arrow_forward