A 1B parameter decoder-only Transformer model trained on code using a causal-masked objective, which allows inserting/infilling code as well as standard left-to-right generation.

The model was trained on public open-source repositories with a permissive, non-copyleft, license (Apache 2.0, MIT, BSD-2 or BSD-3) from GitHub and GitLab, as well as StackOverflow. Repositories primarily contained Python and JavaScript, but also include code from 28 languages, as well as StackOverflow.

For more information, see our:

A larger, 6B, parameter model is also available at facebook/incoder-6B.

pytorch, tokenizers, and transformers. Our model requires HF's tokenizers >= 0.12.1, due to changes in the pretokenizer.

pip install torch
pip install "tokenizers>=0.12.1"
pip install transformers

See https://github.com/dpfried/incoder for example code.

(Note: the incoder-1B and incoder-6B tokenizers are identical, so 'facebook/incoder-6B' could also be used.)

When calling tokenizer.decode, it's important to pass clean_up_tokenization_spaces=False to avoid removing spaces after punctuation. For example:

(Note: encoding prepends the <|endoftext|> token, as this marks the start of a document to our model. This token can be removed from the decoded output by passing skip_special_tokens=True to tokenizer.decode.)

CC-BY-NC 4.0

The model was developed by Daniel Fried, Armen Aghajanyan, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Wen-tau Yih, Luke Zettlemoyer and Mike Lewis.

Thanks to Lucile Saulnier, Leandro von Werra, Nicolas Patry, Suraj Patil, Omar Sanseviero, and others at HuggingFace for help with the model release, and to Naman Goyal and Stephen Roller for the code our demo was based on!