anthonypjshaw/SWE-bench_Complex

SWE-bench Complex

A contamination-free, complexity-focused evaluation set for AI coding agents.

SWE-bench Complex is a curated dataset of 115 real-world GitHub issues from major Python open-source projects, designed specifically for studying code complexity in AI-generated patches. All tasks were merged between January–March 2026, guaranteeing they postdate the training cutoff of current frontier models.

Task-Text Judge Comparison

GPT-5.4 reviewed solver-visible text across Complex (115), Lite (300), Verified (500), and Pro (731) using independent 0-3 scales for clarity concern, internal inconsistency and visible repair disclosure.

All 1,646 task memberships are included, including 30 reviews with quotation flags and three unresolved disclosure ratings. These are uncalibrated model judgments, not human-validated quality scores. Disclosure measures visible guidance, not contamination, correctness or a quality penalty. A recent merge date alone cannot guarantee an unseen task.

The judge method, full rubric and comparisons explain inputs, repetitions, deterministic response selection, exact quotation checks, denominators and limitations. Separate PNGs are available for clarity, consistency, and repair disclosure, alongside vector PDFs, the exact prompt and schema, aggregate counts, pinned dataset revisions and artifact hashes.

This documentation release does not change the 115-task dataset. The judge study uses frozen dataset revisions recorded in its manifest; it does not use a moving repository HEAD.

Why SWE-bench Complex?

Existing benchmarks like SWE-bench Verified suffer from two problems for complexity research:

1. Data Contamination

Over 94% of SWE-bench issues predate current LLM training cutoffs. Aleithan et al. found that 32.67% of successful patches involve "cheating" through solution leakage, and resolution rates dropped from 12.47% to 3.97% when leaked instances were filtered out (SWE-bench+, 2024).

All SWE-bench Complex instances postdate the training cutoffs of:

Model	Provider	Training Cutoff	Gap
Claude Opus 4.6	Anthropic	Oct 2025	3+ months
GPT-5.3-Codex	OpenAI	Sep 2025	4+ months
GPT-5.4	OpenAI	Nov 2025	2+ months
Gemini 3.1 Pro	Google	Oct 2025	3+ months

2. Trivial Patches

SWE-bench Verified has a median patch size of just 7 changed lines — 44.6% of tasks require only 1–5 lines. These trivial patches yield near-zero complexity deltas, reducing statistical power for quality studies.

SWE-bench Complex targets substantive patches with a median of 45 changed lines — 6.4× larger than SWE-bench Verified.

Dataset Comparison

Characteristic	SWE-bench Verified	SWE-bench Complex
Tasks	500	115
Repositories	12	8
Median changed lines	7	45
Mean changed lines	14.3	74.4
Mean Python files changed	1.2	3.9
Human ΔCC (mean)	+1.14	+4.06
Human ΔLLOC (mean)	+2.77	+19.08
Human ΔMI (mean)	−0.230	−0.417
Human ΔCogC (mean)	N/A	+3.63
Post-training-cutoff	<6%	100%

Complexity metrics measured using Wily v2:

• ΔCC: Cyclomatic Complexity change (McCabe, 1976)
• ΔLLOC: Logical Lines of Code change
• ΔMI: Maintainability Index change (Oman & Hagemeister, 1992)
• ΔCogC: Cognitive Complexity change (Campbell, 2018)

Repository Distribution

Repository	Instances
django/django	37
astropy/astropy	22
pydata/xarray	15
scikit-learn/scikit-learn	14
matplotlib/matplotlib	9
pylint-dev/pylint	9
sympy/sympy	8
pallets/flask	1

Selection Criteria

Instances were collected from merged pull requests in the SWE-bench ecosystem repositories with the following filters:

1. Date range: Merged January 1 – March 10, 2026 (post-training-cutoff)
2. Issue linkage: PR explicitly references a GitHub issue via "fixes #N" or equivalent
3. Test coverage: PR includes both implementation and test changes to Python files
4. Minimum complexity: Implementation patch modifies ≥4 changed lines
5. Python files: Only .py file changes retained
6. Manual review: Each candidate reviewed for solvability — documentation-only changes, large-scale refactors (>300 lines or >10 files), and tasks requiring external domain knowledge were excluded

From 1,043 scraped PRs → 712 with issue references → 224 after automated filters → 119 after manual review → 115 with validated four-state test contracts.

Schema

Each instance contains:

Field	Type	Description
instance_id	string	Unique identifier ({owner}__{repo}-{pr_number})
repo	string	GitHub repository (owner/repo)
base_commit	string	Parent commit SHA
problem_statement	string	GitHub issue text (title + body)
test_patch	string	Unified diff of test-file changes
human_patch	string	Unified diff of implementation-file changes
pr_number	int	Pull request number
pr_url	string	Pull request URL
pr_merged_at	string	Merge timestamp (ISO 8601)
issue_number	int	Referenced issue number
issue_url	string	Issue URL
human_changed_lines	int	Total changed lines in the human patch
FAIL_TO_PASS	string	JSON array of test IDs that must go FAIL→PASS
PASS_TO_PASS	string	JSON array of test IDs that must remain PASS

SWE-bench Compatibility

SWE-bench Complex uses the same schema as SWE-bench Verified and can be evaluated using the standard SWE-bench harness:

python -m swebench.harness.run_evaluation \
    -d anthonypjshaw/SWE-bench_Complex \
    -s test \
    -p predictions.jsonl \
    -id my_run \
    --max_workers 4

Test construction and reproducibility

The test-construction report documents candidate screening, environment recovery, the B0/B1/G1/G2 protocol, classification rules, explicit test-patch repairs, and the four excluded candidates.

The reproducibility/ directory contains the generated ARM64 Docker build contexts for all 115 accepted instances: one base layer, 72 deduplicated environment layers, and one instance layer per task. It also contains a manifest, SHA-256 checksums, a layered image builder, the frozen environment evidence, contract audit, and the exact runner and override code.

python reproducibility/build_images.py --instance django__django-20766

Usage

from datasets import load_dataset

dataset = load_dataset("anthonypjshaw/SWE-bench_Complex", split="test")
print(f"Tasks: {len(dataset)}")
print(f"Repos: {len(set(dataset['repo']))}")

Citation

If you use SWE-bench Complex in your research, please cite:

@inproceedings{Shaw2026SWEbenchComplex,
  author    = {Shaw, Anthony},
  title     = {Beyond the Benchmark: A Contamination-Free Study of {AI} Code Complexity Across Four Frontier Models},
  booktitle = {Proceedings of the IEEE International Conference on Software Engineering (SSE)},
  year      = {2026},
}

License

MIT License. The dataset contains references to publicly available open-source code under their respective licenses.