Daniel Fried*, Armen Aghajanyan*, Jessy Lin, Sida Wang, Eric Wallace, Freda Shi, Ruiqi Zhong, Wen-tau Yih, Luke Zettlemoyer, and Mike Lewis

ICLR 2023

This repository hosts example code showing how to use the model using HuggingFace's transformers library. Code to replicate the evaluation results in our paper (in Fairseq, which we used to train the model) is coming soon!

See our project site for more information, or our paper, or examples.

You can obtain the models from HuggingFace's hub:

• 6.7B parameter model: facebook/incoder-6B
• 1.3B parameter model: facebook/incoder-1B

We use a custom tokenizer, which you can load from either "facebook/incoder-1B" or "facebook/incoder-6B" (they are identical). The model was trained with padding on the left-hand side of inputs, using a <pad> token which has ID 1.

tokenizer = AutoTokenizer.from_pretrained("facebook/incoder-1B") # or "facebook/incoder-6B"
tokenizer.pad_token = "<pad>"
tokenizer.padding_side = "left"

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

tokenizer.decode(tokenizer.encode("from ."), clean_up_tokenization_spaces=False)

(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.)

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'
pip install transformers

See example_usage.py for a demo script showing how to use the infilling capability of the model. Set BIG_MODEL = True in the script to use the 6.7B parameter model; the 1.3B will be used otherwise.

For an example of batched generation, see example_batched_usage.py.

See our paper for research details on the method, training data, models, and experimental results.

See a demo of the 6.7B model on HF Spaces. [currently not working, apologies!]

CC-BY-NC 4.0

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!