Saved in:
Bibliographic Details
Main Authors: Sapronov, Maksim, Glukhov, Evgeniy
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.13697
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911212110872576
author Sapronov, Maksim
Glukhov, Evgeniy
author_facet Sapronov, Maksim
Glukhov, Evgeniy
contents Repository-level pretraining is commonly used to enable large language models for code to leverage codebase-wide context. This enhances their ability to generate accurate and context-aware code completions. In this work, we investigate how different repository-processing strategies affect in-context learning in OpenCoder, a 1.5B-parameter model. We extend its context window from 4,096 to 16,384 tokens by training on additional 1B tokens of curated repository-level data. Despite relying on a smaller dataset than competing models (which often use hundreds of billions of tokens), our model achieves comparable performance on the Long Code Arena benchmark. We find that various repository-processing techniques yield similarly strong results, with the primary gain coming from adapting to a new rotary positional embedding (RoPE) scaling parameter. Finally, we show that a simpler file-level training approach at the original sequence length remains highly effective, opening up repository-level code completion research to settings with more constrained data and compute resources.
format Preprint
id arxiv_https___arxiv_org_abs_2510_13697
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On Pretraining for Project-Level Code Completion
Sapronov, Maksim
Glukhov, Evgeniy
Software Engineering
Machine Learning
Repository-level pretraining is commonly used to enable large language models for code to leverage codebase-wide context. This enhances their ability to generate accurate and context-aware code completions. In this work, we investigate how different repository-processing strategies affect in-context learning in OpenCoder, a 1.5B-parameter model. We extend its context window from 4,096 to 16,384 tokens by training on additional 1B tokens of curated repository-level data. Despite relying on a smaller dataset than competing models (which often use hundreds of billions of tokens), our model achieves comparable performance on the Long Code Arena benchmark. We find that various repository-processing techniques yield similarly strong results, with the primary gain coming from adapting to a new rotary positional embedding (RoPE) scaling parameter. Finally, we show that a simpler file-level training approach at the original sequence length remains highly effective, opening up repository-level code completion research to settings with more constrained data and compute resources.
title On Pretraining for Project-Level Code Completion
topic Software Engineering
Machine Learning
url https://arxiv.org/abs/2510.13697