abap-bench 4.0.0

ABAP-Bench: A Comprehensive Benchmark for Evaluating LLM Understanding of SAP ABAP and S/4HANA Modernization

ABAP-Bench

A professional AI benchmark for evaluating LLM understanding of SAP ABAP and S/4HANA modernization.

专业评估大语言模型对 SAP ABAP 及 S/4HANA 现代化理解能力的基准测试

---

Overview / 概述

ABAP-Bench evaluates large language models on real-world SAP enterprise software tasks: migrating legacy ABAP code to S/4HANA APIs, detecting defects in business-critical programs, rewriting classic reports as Fiori applications, and more.

Enterprise software modernization is a $50B+ market. As organizations migrate from ECC to S/4HANA, LLMs are increasingly used to assist developers — yet no standardized benchmark existed to measure their true capability in this domain. ABAP-Bench fills that gap.

Why it matters / 为什么重要：

• SAP systems manage >70% of global transaction revenue; migration errors have real financial consequences
• ABAP is a niche language with sparse training data; general coding benchmarks do not capture domain-specific knowledge
• S/4HANA introduces breaking API changes (BAPI → I_Journal APIs, classical ALV → CL_SALV, etc.) that require precise understanding
• China-specific regulatory and localization requirements (Golden Tax, ChinaTax, PIPL) demand dedicated evaluation coverage

---

Leaderboard (v4.0) / 排行榜

Last updated: 2026-04-03 | Scoring: 4-layer (Rubric 30% + Quality 20% + Semantic 20% + Judge 30%)

Rank	Model	Score (/100)	Migration	Defects	Rewriting	China	Risk	Security	Architecture	Performance	Ecosystem
1	Qwen3 235B	75	—	—	—	—	—	—	—	—	—
2	Grok 4	75	—	—	—	—	—	—	—	—	—
3	GLM-5.1	74	—	—	—	—	—	—	—	—	—
4	DeepSeek R1 (0528)	73	—	—	—	—	—	—	—	—	—
5	MiMo-V2-Pro	68	—	—	—	—	—	—	—	—	—
6	MiniMax M2.7	66	—	—	—	—	—	—	—	—	—

Scores above are from v3.2 (30 tasks, rubric-only scoring). Full v4.0 re-evaluation with 60 tasks and 4-layer scoring is in progress.

Submit your model via Issues.

---

Quick Start / 快速开始

# Clone and install
git clone https://github.com/abap-bench/abap-bench
cd abap-bench
pip install -e .

# Configure API keys
cp .env.example .env
# Edit .env with your keys

# Run full benchmark (all models in configs/models.yaml)
python -m src.run_benchmark

# Run single model
python -m src.run_benchmark --model glm-5.1

# Score a response (3-layer, no API needed)
python -m src.evaluate_v2 --task T01 --response "I_JournalEntry ACDOCA..." --breakdown

# Score with LLM-as-Judge (4-layer, needs API key)
python -m src.evaluate_v2 --task T01 --response "..." --with-judge --judge-model glm-4-plus --judge-backend zhipuai --breakdown

# Run judge on a full result file
python -m src.judge batch results/v4.0/glm-5.1.json --output results/judge/glm-5.1.json

---

Benchmark Design / 基准设计

9 Evaluation Dimensions / 9 个评估维度

┌──────────────────────────────────────────────────────────────────────────┐
│                           ABAP-Bench v4.0                                │
│                        60 Tasks × 20 pts = 1200                          │
├───────────────┬───────────────┬───────────────┬──────────────────────────┤
│  D1 Migration │  D2 Defects   │  D3 Rewriting │  D4 China Compliance     │
│  T01,T09,T10  │  T02,T11,T12  │  T03,T13,T14  │  T04,T15,T16             │
│  T31,T32,T33  │  T34,T35,T36  │  T37,T38,T39  │  T40,T41,T42             │
├───────────────┼───────────────┼───────────────┼──────────────────────────┤
│  D5 Risk      │  D6 Security  │  D7 Architect │  D8 Performance          │
│  T05,T17,T18  │  T06,T19,T20  │  T07,T21,T22  │  T08,T23,T24             │
│  T43,T44,T45  │  T46,T47,T48  │  T49,T50,T51  │  T52,T53,T54             │
├───────────────┴───────────────┴───────────────┴──────────────────────────┤
│  D9 Modern Ecosystem: T25-T30, T55-T60 (12 tasks)                       │
│  Clean Core · Unit Testing · Fiori · BAdI · LUW · Integration Suite      │
│  Workflow · Output Mgmt · IDoc/ALE · BDC · Change Mgmt · Code Inspector  │
└──────────────────────────────────────────────────────────────────────────┘

#	Dimension	Tasks	Description
D1	Code Migration	6	ECC → S/4HANA API replacement (BKPF→ACDOCA, BAPI→Released API)
D2	Defect Discovery	6	Finding hidden bugs in ABAP code (N+1 queries, scope leaks, silent data loss)
D3	Code Rewriting	6	Modernizing classical ABAP to clean code, RAP, ABAP Cloud
D4	China Compliance	6	Golden Tax, ChinaTax VAT, PIPL privacy, 五险一金 payroll
D5	Migration Risk	6	Change impact analysis, RFC dependency chains, transport risks
D6	Security & Auth	6	Authority checks, SQL injection, authorization trace, transport security
D7	S/4HANA Architecture	6	ACDOCA, CDS views, FI-CO integration, ledger architecture
D8	Performance Engineering	6	SELECT optimization, HANA column store, parallel processing
D9	Modern Ecosystem	12	Clean Core, unit testing, Fiori, BAdI, LUW, IDoc, workflow, BDC

4-Layer Scoring / 四层评分

┌─────────────────────────────────────────────────────────────┐
│  Layer 4: LLM-as-Judge (optional)              30%         │
│  ├── Correctness · Completeness · Specificity              │
│  ├── Structure · Insight (each 1-5, total /25)             │
│  └── Reference-guided via golden answers                   │
├─────────────────────────────────────────────────────────────┤
│  Layer 3: Semantic Similarity                  20% (30%*)  │
│  ├── BM25 text similarity against golden answers           │
│  └── Concept coverage (key_concepts hit rate)              │
├─────────────────────────────────────────────────────────────┤
│  Layer 2: Code & Structure Quality             20% (30%*)  │
│  ├── ABAP syntax checks (for code tasks)                   │
│  └── Answer structure analysis (for knowledge tasks)       │
├─────────────────────────────────────────────────────────────┤
│  Layer 1: Rubric Matching                      30% (40%*)  │
│  ├── Keyword matching (weighted_terms, keyword_group)      │
│  ├── Compound matching (key_term + context_keywords)       │
│  └── Penalty rules (incorrect S/4HANA statements: -1~-3)  │
└─────────────────────────────────────────────────────────────┘
  * Weights in parentheses: 3-layer mode (Layer 4 disabled)

3-layer mode (default, no API needed): Rubric 40% + Quality 30% + Semantic 30% 4-layer mode (--with-judge): Rubric 30% + Quality 20% + Semantic 20% + Judge 30%

---

Project Structure / 项目结构

ABAP-Bench/
├── README.md                    # This file
├── pyproject.toml               # Python packaging (pip install -e .)
├── benchmark_card.yaml          # HuggingFace Dataset Card format
├── CITATION.cff                 # Citation metadata
├── CHANGELOG.md                 # Version history
├── LICENSE                      # Apache-2.0
├── .env.example                 # API key template
│
├── data/
│   ├── tasks.jsonl              # 60 task definitions (JSONL)
│   ├── dimensions.json          # 9 dimensions metadata
│   ├── rubrics/                 # 60 scoring rubric JSONs (T01-T60)
│   ├── golden/                  # 60 golden reference answers (T01-T60)
│   └── test_code/               # ABAP code samples for code-review tasks
│       ├── zvat_invoice_process.abap
│       ├── zhr_salary_calc.abap
│       └── zdyn_query.abap
│
├── src/
│   ├── __init__.py              # Package init (version: 4.0.0)
│   ├── run_benchmark.py         # Main runner: load tasks → call LLM → score → save
│   ├── evaluate.py              # Scoring engine v1 (T01-T30, rubric-only)
│   ├── evaluate_v2.py           # Scoring engine v2 (4-layer, all 60 tasks)
│   ├── judge.py                 # LLM-as-Judge module (Layer 4)
│   └── models.py                # Multi-backend LLM client (zero external deps)
│
├── configs/
│   └── models.yaml              # Model registry (7 models, 4 backends)
│
├── results/
│   ├── schema.json              # Result file JSON Schema
│   └── v4.0/                    # Per-model evaluation results
│
├── scripts/
│   ├── validate_rubrics.py      # Data integrity validation
│   └── migrate_from_legacy.py   # v3.2 → v4.0 migration helper
│
├── tests/                       # Unit & integration tests
│   ├── test_evaluate.py
│   ├── test_data_integrity.py
│   └── test_judge.py
│
└── docs/
    ├── DECONTAMINATION.md       # Data provenance & contamination statement
    └── IMPLEMENTATION_PLAN.md   # Development roadmap (P0-P5)

---

Adding New Tasks / 添加新任务

1. Append a JSON line to data/tasks.jsonl:

{"task_id":"T61","title":"New Task","dimension":"Code Migration Knowledge","max_score":20,"prompt_template":"...","requires_test_code":false,"version":"4.1"}

2. Create rubric: data/rubrics/T61.json
3. Create golden answer: data/golden/T61.json
4. Update data/dimensions.json to include T61
5. Validate: python scripts/validate_rubrics.py
6. Test: python -m src.evaluate_v2 --task T61 --response "..." --breakdown

---

Known Limitations / 已知局限

• No code execution: ABAP requires licensed SAP systems; scoring relies on static analysis + LLM-as-Judge instead of unit tests
• 60 tasks: Below the 100+ statistical significance threshold of top benchmarks (SWE-bench: 2294, BigCodeBench: 1140)
• No human correlation study: Inter-annotator agreement not yet measured (planned: Spearman ρ target > 0.85)
• Primarily Chinese prompts: May disadvantage models weaker in Chinese language understanding

See IMPLEMENTATION_PLAN.md for the full future roadmap.

---

Citation / 引用

@misc{abapbench2026,
  title        = {ABAP-Bench: A Benchmark for Evaluating LLM Understanding of SAP ABAP and S/4HANA Modernization},
  author       = {ABAP-Bench Contributors},
  year         = {2026},
  version      = {4.0},
  howpublished = {\url{https://github.com/abap-bench/abap-bench}},
  note         = {60 tasks, 9 dimensions, 4-layer scoring}
}

---

License / 许可证

This project is licensed under the Apache License 2.0. See LICENSE for details.

Benchmark task prompts and rubrics are released under Apache-2.0. Golden reference answers (data/golden/) are provided for evaluation use only and should NOT be included in LLM training data. Model responses collected during evaluation remain the property of their respective model providers.