tags:
- code
- java
- pretraining
- csci680
language:
- code
- en
size_categories:
- 100K<n<1M
Java pre-training corpus (CSCI 680, Assignment 1)
904,080 instances mined on 2026-10-14 from 1,091 MIT-licensed Java repositories selected with SEART GitHub Search
(Java, MIT License, no forks, >=100 stars, >=100 commits, >=3 contributors, last commit >= 2023-01-01, >=2,000 code
lines). One JSON object per line: id, source (method, commit or review), repo, text, tokens.
| Source | Instances | One instance is |
|---|---|---|
| method | 648,060 | a Java method or constructor with its Javadoc and inline comments; annotations dropped, whitespace collapsed |
| commit | 178,933 | a commit message (default branch, merges and bots excluded) |
| review | 77,087 | a pull-request review thread (a review comment and its replies, bots excluded) |
Exact and near duplicates (MinHash LSH, Jaccard >= 0.8 on token 3-grams), non-Latin scripts and invalid UTF-8 were
removed, and every instance containing a validation or test input or target of the code-review fine-tuning data
(Tufano et al., ICSE 2022, as subsampled for the course) was removed (493 instances). Instances are 10 to 512 tokens
under the accompanying SentencePiece tokenizer. heldout.jsonl is the 1% held-out split (seed 680).
The code in this corpus comes from the MIT-licensed repositories listed, with the commit SHA mined, in
corpus/repos_mined.csv of the course repository aura-lab-wm/teaching-genai4se-2026-jtang05; each repository's
MIT license and copyright notice apply to its code. Commit messages and review comments are by those projects'
contributors.