---
license: mit
task_categories:
  - text-generation
  - other
language:
  - en
  - code
tags:
  - unit-test-generation
  - repository-level
  - code-generation
  - benchmark
pretty_name: XRepoTest
---

# XRepoTest: Multilingual Repository-Level Unit Test Generation Benchmark

The data release for the XRepoTest benchmark (EMNLP 2026 Main).

Each task item is a **function extracted from a real open-source repository**. The
goal is to generate a unit test for that function using its surrounding repository
context. Generated tests are executed and scored inside Docker on the real codebase.

## Structure

| Folder | Contents | Pipelines |
|--------|----------|-----------|
| `data/base` | per-language task items `{lang}_functions.jsonl` | `standard`, `file_context` |
| `data/lsp`  | LSP-enriched task items `{lang}_functions_enriched.jsonl` | `lsp_context` |
| `data/rag`  | retrieval-enriched items (`ws<window>_ss<step>_k<top_k>`, bm25/unixcoder) | `rag_bm25`, `rag_dense` |

## Languages

Go, Rust, Julia, PHP, Ruby — one folder per language. See the repository
[README](https://github.com/) for pipeline usage.

## Record fields (base)

- `task_id`, `function_name`, `file_path`
- `focal_code` — the function under test
- `file_content` — full source of the containing file
- `function_component` / `metadata` — parsed structure and provenance

## Reproducibility

`repo_data/` (the raw crawled source repositories) is **not** published here. The
per-language Docker evaluation images bake the corpus in; dataset construction is
reproducible from the public source repositories listed in each record's metadata.

## License

Dataset release under the MIT license. Individual source repositories retain their
own licenses; see provenance metadata in each record.