SantaCoder

Play with the model on the SantaCoder Space Demo.

Model Summary

The SantaCoder models are a series of 1.1B parameter models trained on the Python, Java, and JavaScript subset of The Stack (v1.1) (which excluded opt-out requests). The main model uses Multi Query Attention, was trained using near-deduplication and comment-to-code ratio as filtering criteria and using the Fill-in-the-Middle objective. In addition there are several models that were trained on datasets with different filter parameters and with architecture and objective variations.

Model Architecture Objective Filtering
mha MHA AR + FIM Base
no-fim MQA AR Base
fim MQA AR + FIM Base
stars MQA AR + FIM GitHub stars
fertility MQA AR + FIM Tokenizer fertility
comments MQA AR + FIM Comment-to-code ratio
dedup-alt MQA AR + FIM Stronger near-deduplication
final MQA AR + FIM Stronger near-deduplication and comment-to-code ratio

The final model is the best performing model and was trained twice as long (236B tokens) as the others. This checkpoint is the default model and available on the main branch. All other checkpoints are on separate branches with according names.

Use

Intended use

The model was trained on GitHub code. As such it is not an instruction model and commands like "Write a function that computes the square root." do not work well. You should phrase commands like they occur in source code such as comments (e.g. # the following function computes the sqrt) or write a function signature and docstring and let the model complete the function body.

Feel free to share your generations in the Community tab!

How to use

Generation

# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigcode/santacoder"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, trust_remote_code=True).to(device)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))


Fill-in-the-middle

Fill-in-the-middle uses special tokens to identify the prefix/middle/suffic part of the input and output:

input_text = "<fim-prefix>def print_hello_world():\n    <fim-suffix>\n    print('Hello world!')<fim-middle>"
inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))


We upload the checkpoint of each experiment to a separate branch as well as the intermediate checkpoints as commits on the branches. You can load them with the revision flag:

model = AutoModelForCausalLM.from_pretrained(
"bigcode/santacoder",
revision="no-fim", # name of branch or commit hash
trust_remote_code=True
)


The pretraining dataset of the model was filtered for permissive licenses only. Nevertheless, the model can generate source code verbatim from the dataset. The code's license might require attribution and/or other specific requirements that must be respected. We provide a search index that let's you search through the pretraining data to identify where generated code came from and apply the proper attribution to your code.

Limitations

The model has been trained on source code in Python, Java, and JavaScript. The predominant language in source is English although other languages are also present. As such the model is capable to generate code snippets provided some context but the generated code is not guaranteed to work as intended. It can be inefficient, contain bugs or exploits.

Training

Model

• Architecture: GPT-2 model with multi-query attention and Fill-in-the-Middle objective
• Pretraining steps: 600K
• Pretraining tokens: 236 billion
• Precision: float16

Hardware

• GPUs: 96 Tesla V100
• Training time: 6.2 days
• Total FLOPS: 2.1 x 10e21

Software

The model is licenses under the CodeML Open RAIL-M v0.1 license. You can find the full license here.

Citation

@article{allal2023santacoder,
title={SantaCoder: don't reach for the stars!},
author={Allal, Loubna Ben and Li, Raymond and Kocetkov, Denis and Mou, Chenghao and Akiki, Christopher and Ferrandis, Carlos Munoz and Muennighoff, Niklas and Mishra, Mayank and Gu, Alex and Dey, Manan and others},
journal={arXiv preprint arXiv:2301.03988},
year={2023}
}