---
pretty_name: Code QA Unified
license: other
license_name: mixed-upstream-terms
license_link: https://huggingface.co/datasets/takara-ai/code-qa-unified/blob/main/LICENSES.md
language:
- en
- code
- ru
- ja
- tr
task_categories:
- text-retrieval
- question-answering
tags:
- code-search
- code-qa
- deduplicated
- provenance
- license-filtered
size_categories:
- 10M<n<100M
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path:
    - code_search/train/*.parquet
    - other_qa/train/*.parquet
- config_name: code_search
  data_files:
  - split: train
    path: code_search/train/*.parquet
- config_name: other_qa
  data_files:
  - split: train
    path: other_qa/train/*.parquet
---

# Code QA Unified

**88,234,786 unique positive query–answer pairs**, including **86,686,806 text-to-code retrieval pairs**.
Curated by **Jordan Legg**. Maintained by **Takara AI**.

```python
from datasets import load_dataset
ds = load_dataset("takara-ai/code-qa-unified", "code_search", split="train", streaming=True)
```

Use `streaming=True` to iterate without loading the full dataset into memory.

## Contents

| Configuration | Pairs | Content |
|---|---:|---|
| `default` | 88,234,786 | All positive QA pairs |
| `code_search` | 86,686,806 | Natural-language queries paired with code |
| `other_qa` | 1,547,980 | Other positive QA tasks |

Each configuration has one `train` split. The dataset combines existing sources under a common schema with exact-pair deduplication, pinned provenance and evaluation-overlap filtering.

## Curation and provenance

Rows preserve question and answer text, source repository/configuration/split/revision, original IDs, language, license, repository/path/URL, JSON metadata and a binary SHA-256 `content_id`. Repository and path coverage varies by source.

Exact duplicates are identified by question and answer hashes plus ordered field types, with line endings normalized. Multiple questions per code and multiple positive answers per question remain separate pairs. Filtering preserves retained field values.

Included sources declare MIT, Apache-2.0, CC BY 4.0 or CC BY-SA 4.0. Sources with unspecified, custom, noncommercial or no-derivatives terms are excluded. Rows with other or unrecognised license identifiers are omitted, including unknown original-code licenses in the Shuu collections. This filter excluded **7,639,316 pairs**. [The publication audit](build/publication-audit.json) records counts and decisions.

## Licensing and attribution

Upstream creators retain credit for the code, queries, annotations, translations and source datasets. [ATTRIBUTIONS.md](ATTRIBUTIONS.md) lists pinned sources; [CITATIONS.md](CITATIONS.md) contains their citations. Source cards and available notices are preserved in `upstream-cards/` and `upstream-notices/`.

[LICENSES.md](LICENSES.md) records source-specific terms. Original code may carry separate repository licenses. Required copyright notices, license texts and change disclosures must be retained; CC BY-SA material remains subject to applicable share-alike conditions. The collection does not relicense upstream material.

## Evaluation overlap and limitations

Pairs were checked against **1,538 evaluation queries** and **2,059 positive documents** from `takarajordan/code-search-eval-audited` at revision `50b8424c85cf929d3075dbbaa2bc34b4cd8c589a`. Exact matches and lexical shingle Jaccard ≥0.90 matches were excluded, checking both fields in either direction.

The overlap checks cover that evaluation companion only. Semantic near-duplicates and repository-level overlap may remain. Positive labels have not been independently verified for relevance or execution correctness.

Credential cleanup excluded 118 credential-bearing or dependent provenance records from the source build. Those checks covered two reported upstream keys.

## Files and audit

[files.json](files.json) lists public shards, row counts, sizes and SHA-256 checksums. [build/publication-audit.json](build/publication-audit.json) records the pinned input and filtering policy.