---
license: apache-2.0
task_categories:
- text-classification
language:
- en
tags:
- java
- compile-errors
- codenet
- education
pretty_name: CodeNet Java compile errors v2
size_categories:
- 10K<n<100K
---

# CodeNet Java compile errors v2

This is a **new, Java-only corpus** rebuilt from
[`iNeil77/CodeNet`](https://huggingface.co/datasets/iNeil77/CodeNet) (the Java
split, `status == "Compile Error"`). It is deliberately separate from
[`criyle/codenet-compile-errors`](https://huggingface.co/datasets/criyle/codenet-compile-errors),
which used the old `Main.java` convention and did not carry user/problem/date
metadata.

## What is different

Each submission is written to the file name javac asks for — the single public
top-level class it declares, or `Main.java` when there is no such class or more
than one. Submissions that compile cleanly under that name are **file-name
artefacts**, not student errors; they are kept in `compiled_all.parquet` but
excluded from this dataset.

The source-name fix is why this corpus is smaller than the raw Compile Error
set: `18,465` of
`53,023` submissions compiled cleanly once the file was named
after the class. `34,555` still fail and are
the rows of this dataset.

## Metadata for session grouping

Unlike the old dataset, each row keeps:

| Column | Meaning |
|---|---|
| `row_id` | stable `v2_<sid>` id |
| `sid`, `submission_id` | CodeNet submission id |
| `pid`, `problem_id` | CodeNet problem id |
| `uid`, `user_id` | CodeNet user id |
| `date`, `date_iso` | CodeNet submission time |
| `source_name` | file name used for the compile (`Foo.java` / `Main.java`) |
| `code`, `compilemessage` | source and javac stderr |
| `session_id` | `(uid, pid)` 2-hour session |
| `session_6h` | 6-hour variant |

`session_id` is a contiguous run of submissions by one user on one problem,
with a new session when the gap exceeds 2 hours. It is computed on the **full**
Java split (accepted submissions included), so a user's accepted submission
still closes the session boundary.

## Labels

`category` is intentionally empty and `label_status` is `unlabelled`. The v2.1
taxonomy (`categories.v2.yaml`) is included, but labelling is a separate LLM
pass with its own calibration gates. **No v1 labels are mixed into the `category`
column.**

## Provenance

- source: `iNeil77/CodeNet`, Java split
- filter: `status == 'Compile Error'`
- sandbox: `http://10.0.0.162:5050`
- javac: `OpenJDK 21.0.12.1`
- session gap: `2 h` (`6 h` alternative)
- built at: `2026-10-20T02:28:07.997126+00:00`

## Files

- `train` — the still-failing submissions (`corpus_v2.parquet`)
- `compiled_all.parquet` — every compile attempt, including file-name artefacts
- `session_map_all.parquet` — session map over the full Java split
- `categories.v2.yaml` — active v2.1 taxonomy
- `source_manifest.json` — counts, hashes and build parameters

## Licence

CodeNet is Apache-2.0. This derived corpus is released under the same terms.