---
task_categories:
  - text-classification
language:
  - en
tags:
  - github-issues
  - issue-classification
  - setfit
  - nlbse
  - arxiv:2209.11055
pretty_name: GitHub Issues 4-Class (bug/feature/question/support)
---

# GitHub Issues 4-Class

A balanced 4-class dataset for GitHub issue classification: `bug` / `feature` / `question` / `support`.

## Splits

| Split | Rows | Per class |
|---|---|---|
| train | 1,997 | ~500 each (NLBSE rows failing a 15-char minimum-text filter dropped) |
| test | 1,997 | ~500 each |

## Columns

- `text`: issue title + `\n\n` + body, whitespace-normalized, capped at 4,000 chars (models truncate to 256 tokens)
- `label`: one of `bug`, `feature`, `question`, `support`
- `repo`: source GitHub repository
- `source`: `nlbse2024` or `victor-betus`

## Provenance

- **bug / feature / question**: the NLBSE'24 issue-report-classification benchmark (3,000 balanced issues from facebook/react, tensorflow/tensorflow, microsoft/vscode, bitcoin/bitcoin, opencv/opencv), raw CSVs from `github.com/nlbse2024/issue-report-classification`. Labels are maintainer-assigned GitHub labels mapped by the benchmark's synonym table.
- **support**: issues labeled `support`-family (`support`, `site-support-request`, `help wanted`, `type: support`, ...) filtered from [`victor-betus/github-issues-dataset`](https://huggingface.co/datasets/victor-betus/github-issues-dataset) (114k issues, top-100 GitHub repos). Issues that also carry a conflicting class label (bug/enhancement/feature/question) were excluded, and all rows were deduplicated against the NLBSE benchmark by (repo, normalized title) — zero overlap.

## Known caveats

- GitHub label conventions vary per repo; cross-project label inconsistency is the dominant error source in this task (Izadi et al., MSR 2024).
- The `question` class is historically the hardest (F1 ~0.70 in our runs); `support` is easiest.
- `help wanted` is treated as a support-request synonym following GitHub convention in several large repos; it can also denote "contributors wanted" in others — some label noise in the support class is possible.

## Baseline results (this dataset, test split)

| Model | macro-F1 (4-class) |
|---|---|
| SetFit all-MiniLM-L6-v2 (contrastive + logistic head) | 0.807 |
| DistilBERT fine-tune (lr 2e-5, 4 epochs, 256 tokens) | see model card |

## References

- Colavito, Lanubile, Novielli, "Few-Shot Learning for Issue Report Classification", NLBSE 2023
- Kallis et al., "The NLBSE'24 Tool Competition" (issue report classification track)
- Tunstall et al., "Efficient Few-Shot Learning Without Prompts" (SetFit), arXiv:2209.11055