NLPForUA/multilingual-repo-qa

UPD: Release in progress, stay tuned.

MultiRepoQA

MultiRepoQA is a multilingual benchmark for question answering over complete software repositories. It contains 783 validated canonical questions across 30 open-source repositories, with aligned English, German, and Ukrainian versions (2,349 language-specific examples). Questions cover local implementation details, cross-file behavior, repository-wide flows, and maintenance impact.

The benchmark was introduced in Multilingual Repository-Level Code Question Answering Benchmark, published in the proceedings of MIT-2026, pp. 358–361 (DOI: 10.5281/zenodo.20793311). That short paper describes an earlier full-context evaluation stage. This release applies a later, stricter union validity filter and additionally includes tree- and SiMAL-based retrieval experiments, so its retained counts differ from the paper.

The benchmark also releases predictions and LLM-as-a-judge annotations for five model/strategy combinations:

Configuration	Answer model	Context strategy	Rows
questions	—	Gold questions, references, and evidence	2,349
repositories	—	Repository metadata and pinned revisions	30
simal_schemas	gpt-5.4-2026-03-05	Exact compact schemas used by SiMAL runs	30
gpt-5.4_full-context	gpt-5.4-2026-03-05	Complete serialized repository	2,349
gpt-5.4_2stage-tree	gpt-5.4-2026-03-05	Repository tree retrieval, then selected files	2,349
gpt-5.4_2stage-simal	gpt-5.4-2026-03-05	SiMAL retrieval, then selected files when needed	2,349
gpt-oss-20b_2stage-tree	openai/gpt-oss-20b	Repository tree retrieval, then selected files	2,349
gpt-oss-20b_2stage-simal	openai/gpt-oss-20b	SiMAL retrieval, then selected files when needed	2,349
traces_gpt-5.4_2stage-tree	gpt-5.4-2026-03-05	Complete saved tree requests and outputs	2,349
traces_gpt-5.4_2stage-simal	gpt-5.4-2026-03-05	Complete saved SiMAL requests and outputs	2,349
traces_gpt-oss-20b_2stage-tree	openai/gpt-oss-20b	Complete requests and sanitized provider responses	2,249
traces_gpt-oss-20b_2stage-simal	openai/gpt-oss-20b	Complete requests and sanitized provider responses	2,341

Context strategies

• Full context: serializes all included repository files into a single answering request.
• Two-stage tree: stage 1 sees a repository tree and selects relevant paths; stage 2 answers from the retrieved file contents.
• Two-stage SiMAL: stage 1 sees the repository tree plus a compact SiMAL system schema and may answer directly or request file contents for stage 2.

SiMAL (System internals Modeling and Annotation Language) is a compact, human-readable, error-tolerant, and machine-parseable DSL for software-system context. It unifies static structure and dynamic behavior in one textual schema: architectural views, components, endpoints and contracts, dependencies, cross-component interactions, runtime/deployment metadata, configuration, and links back to source artifacts. Unlike a plain file tree or symbol map, it is intended to retain both high-level system orientation and lower-level implementation relationships. Unlike generic JSON serialization, its grammar is designed to reduce repeated syntactic overhead in LLM prompts. The language has a normalized machine representation, parser, and nested-diagram visualizer. See https://syromiatnikov.me/publications/simal-2026.pdf.

Dataset design

One row per language version

Each language version is a separate row. The canonical_id field is identical for the English, German, and Ukrainian rows, so aligned variants can be grouped without text matching.

The reference answer remains in English for all three question languages. reference_answer_language makes this explicit. Answers should be evaluated semantically rather than by string overlap.

All QA pairs were generated with claude-opus-4-7.

Strict canonical-question validity filter

The original generation produced 900 canonical questions (30 per repository). Then 117 invalid canonical questions removed and 783 left.

The authoritative strategies are full-context, two-stage repository-tree, and two-stage SiMAL.

The evaluation results were produced with gpt-5.5-2026-04-23 at high reasoning effort. All prediction runs used medium reasoning effort.

Excluded-item issue distribution

The table below summarizes the issue annotations attached to the 117 excluded canonical questions. Counts are unique within each severity/issue-type group but are not mutually exclusive: one canonical question can carry more than one issue, so the table must not be summed to obtain the number excluded.

Severity	Issue type	Unique canonical questions
Major	evidence_context_insufficient	2
Major	incomplete_gold_evidence	2
Major	incorrect_reference	7
Major	invalid_question	1
Minor	ambiguous_question	19
Minor	evidence_context_insufficient	3
Minor	incomplete_gold_evidence	50
Minor	incorrect_reference	49
Minor	invalid_question	2
Minor	unsupported_reference	23

Here, incorrect_reference denotes a materially wrong gold answer; unsupported_reference denotes claims not supported by the supplied repository evidence; incomplete_gold_evidence means that the answer may be sound but its gold file set omits required files; and evidence_context_insufficient means the available file contents are insufficient to verify the item. The remaining types identify invalid or materially ambiguous questions.

Context levels

Every canonical question is assigned exactly one context level using the generation rubric below:

Level	Definition	Required gold evidence
L1_local_single_file	Answerable from one file or one very local code region. It is not L1 if following calls across modules or reading two or more files is necessary.	Exactly 1 file
L2_cross_file	Requires several related files, usually within one feature, package, module, or architectural layer.	2–4 files
L3_repository_flow	Requires following a broader feature or runtime flow across modules, layers, packages, services, or runtime components. It tests repository navigation and architectural understanding.	3–6 files
L4_maintenance_impact	Asks about change impact, maintenance risk, test coverage, quality evidence, or what should be inspected before modifying behavior.	3–8 files

The generation target was 25% L1, 35% L2, 25% L3, and 15% L4. Evidence was required to be minimal but sufficient, use exact repository-relative paths, include every file mentioned in the reference answer, and satisfy expected_file_count == len(evidence). After strict filtering, the retained set contains 225 L1, 317 L2, 176 L3, and 65 L4 canonical questions (675, 951, 528, and 195 language rows, respectively).

Results

The extended evaluation below uses the 2,349 language rows retained by the strict filter. Answer is mean answer correctness and Evidence is mean evidence quality. Both are reported on the evaluation rubric's 0–4 scale.

Context strategy	Model	Answer	Evidence
Full context	GPT‑5.4	3.80	3.64
Two-stage tree	GPT‑5.4	3.68	3.60
Two-stage SiMAL	GPT‑5.4	3.68	3.62
Two-stage tree	GPT‑OSS‑20b	3.07	2.99
Two-stage SiMAL	GPT‑OSS‑20b	3.13	3.05

By context level

Each cell is Answer / Evidence; the parenthesized count is the number of evaluated language rows.

Strategy and model	L1 (675)	L2 (951)	L3 (528)	L4 (195)
Full context, GPT‑5.4	3.92 / 4.00	3.83 / 3.62	3.77 / 3.38	3.34 / 3.24
Two-stage tree, GPT‑5.4	3.81 / 3.93	3.76 / 3.62	3.59 / 3.33	3.12 / 3.04
Two-stage SiMAL, GPT‑5.4	3.75 / 3.96	3.72 / 3.56	3.61 / 3.36	3.42 / 3.46
Two-stage tree, GPT‑OSS‑20b	3.27 / 3.51	3.16 / 2.97	2.94 / 2.63	2.31 / 2.25
Two-stage SiMAL, GPT‑OSS‑20b	3.41 / 3.76	3.21 / 2.97	2.93 / 2.58	2.36 / 2.20

By question language

Strategy and model	English Answer / Evidence	German Answer / Evidence	Ukrainian Answer / Evidence
Full context, GPT‑5.4	3.82 / 3.67	3.81 / 3.65	3.78 / 3.61
Two-stage tree, GPT‑5.4	3.69 / 3.61	3.67 / 3.57	3.69 / 3.60
Two-stage SiMAL, GPT‑5.4	3.68 / 3.63	3.69 / 3.61	3.67 / 3.62
Two-stage tree, GPT‑OSS‑20b	3.09 / 2.99	3.08 / 2.98	3.04 / 2.99
Two-stage SiMAL, GPT‑OSS‑20b	3.18 / 3.04	3.12 / 3.03	3.10 / 3.06

SiMAL direct answers and token use

In the SiMAL strategy, stage 1 may answer directly from the tree and schema or request selected source files for stage 2.

Model	Direct from stage 1	Rows	Answer	Evidence
GPT‑5.4	No	1,529	3.77	3.61
GPT‑5.4	Yes	820	3.52	3.63
GPT‑OSS‑20b	No	1,295	3.29	3.12
GPT‑OSS‑20b	Yes	1,046	2.96	2.98

Token totals for the reported two-stage runs are:

Strategy	Model	Input tokens	Generated tokens	Retrieved file-context tokens
Two-stage tree	GPT‑5.4	29.40M	3.99M	22.12M
Two-stage SiMAL	GPT‑5.4	37.40M	3.61M	12.29M
Two-stage SiMAL	GPT‑OSS‑20b	32.77M	3.68M	7.21M

Data fields

questions

• id: unique language-row ID: repo::question_id::language.
• canonical_id: language-independent join key: repo::question_id.
• question_id: repository-local question identifier.
• repo: source repository; URL, pinned Git commit, and license are stored once in the repositories configuration.
• language: en, de, or uk.
• question, question_intent: localized question and generation intent.
• reference_answer, reference_answer_language: English gold answer and its language.
• evidence: list of gold {path, role} objects. Paths are relative to the pinned repository root.
• context_level: L1 local, L2 cross-file, L3 repository flow, or L4 maintenance impact.
• topic, answer_type, difficulty, expected_file_count: question metadata.

repositories

One row per source repository with its URL, pinned revision, detected license, repository kind, main languages/domains, retained-question counts, context serialization statistics, hashes, and selection-stage token statistics.

simal_schemas

One row per repository with the exact SiMAL schema used by the SiMAL prediction runs, character/byte/token counts, and SHA‑256 digest. The schemas were generated with gpt-5.4-2026-03-05. These compact schemas are included because regenerating a model-authored summary is non-deterministic.

Prediction configurations

Every prediction row repeats the question join keys and adds:

• generated answer, evidence_paths, confidence, and cannot_answer;
• textual answer_correctness and evidence_quality labels;
• major_errors, minor_errors, and the concise judge_explanation;
• for two-stage runs: retrieval decision, tree/schema sufficiency where applicable, direct-answer flag, selected/missing paths, retrieved-context size, and per-stage API token usage.

The released judge labels have the following meanings:

answer_correctness	Meaning
correct	Correct and sufficiently complete
mostly_correct	Minor omissions or imprecision
partially_correct	Important information missing or mixed correct/incorrect claims
mostly_incorrect	Only small relevant fragments
incorrect	Incorrect, irrelevant, hallucinated, or unjustified cannot-answer

evidence_quality	Meaning
complete_or_nearly_complete	Covers all or nearly all key evidence
mostly_complete	Covers most evidence with only minor omissions
partially_complete	Some relevant evidence but major files are missing
weak_or_mostly_irrelevant	Weak or mostly irrelevant evidence
missing_or_invalid	No valid evidence or hallucinated paths

Raw prompts and API responses

The four compressed traces_* configurations contain the complete saved requests for the two-stage strategies. Stage 1 contains the repository tree or tree plus SiMAL schema. Stage 2, when executed, contains only the selected file contexts rather than the complete repository. Each row also contains the persisted stage outputs, normalized input/output/total token counts, and final parsed output.

The GPT‑5.4 pipeline did not retain raw provider response envelopes, so its trace rows contain the exact persisted parsed stage outputs. GPT‑OSS provider responses were sanitized: reasoning, raw assistant content, parsed output, and token usage are retained.

GPT‑5.4 has complete trace coverage for all 2,349 retained language rows. The original GPT‑OSS artifact directories lack per-question traces for 100 tree rows and 8 SiMAL rows, although their final aggregate predictions exist. Accordingly, the GPT‑OSS trace configurations contain 2,249 and 2,341 rows.

Usage

from datasets import load_dataset

questions = load_dataset(
    "NLPForUA/multilingual-repo-qa",
    "questions",
    split="test",
)

simal_predictions = load_dataset(
    "NLPForUA/multilingual-repo-qa",
    "gpt-5.4_2stage-simal",
    split="test",
)

uk_questions = questions.filter(lambda row: row["language"] == "uk")

To reproduce repository context, fetch the repository and check out the exact revision recorded in each row:

git clone https://github.com/apache/apisix-dashboard.git
cd apisix-dashboard
git checkout dd641c9b09defc3088cf171fa438d50e44e1f6d3

Repository contents and licensing

Complete repositories are not distributed in this dataset. Exact compact SiMAL schemas are distributed because they are required to reproduce the SiMAL runs, and optional two-stage trace requests embed only the file excerpts retrieved for each question. Gold and predicted evidence in the compact configurations remain relative paths.

CC BY 4.0 applies to the dataset compilation, generated questions, annotations, predictions, and original trace scaffolding. It does not replace upstream rights in source material embedded in stage-2 requests. Source repositories remain under their own licenses; THIRD_PARTY_LICENSES.txt contains the available upstream LICENSE, COPYING, NOTICE, and license-directory texts keyed by repository and pinned commit.

The pinned local license audit is not exclusively MIT/Apache: it includes one AGPL‑3.0 repository (drawdb-io/drawdb), one Unlicense/public-domain repository (vaadin/flow-crm-tutorial), and two dual MIT/Apache repositories. Trace users must comply with the upstream license associated with each embedded excerpt.

Limitations

• Questions and initial reference answers were machine-generated and may retain subtle errors despite strict filtering and human reverification.
• German and Ukrainian questions are aligned with English canonical questions, but the gold reference answers are English only.
• The validity filter is conservative: a single non-valid authoritative observation removes the canonical question from all languages.
• Judge outcomes inherit model-based evaluation biases and should not be treated as human ground truth.
• The benchmark is tied to exact repository commits and does not test knowledge of later versions.
• Compact prediction rows identify evidence paths; optional trace rows may include the selected source contents.
• Publishing predictions can enable result lookup; use a separate hidden test set if contamination-resistant leaderboard evaluation is required.

Citation

Please cite the benchmark paper:

@inproceedings{mit2026,
  author = {Mykyta Syromiatnikov},
  booktitle = {Proceedings of the Sixteenth International Conference of Young Scientists and Students "Modern Information Technology" (MIT-2026)},
  title = {Multilingual Repository-Level Code Question Answering Benchmark},
  year = {2026},
  volume = {1},
  number = {1},
  pages = {358-361},
  doi = {10.5281/zenodo.20793311}
}