Cob2Java-Bench

A benchmarking framework for evaluating LLM-based COBOL-to-Java converters. It runs a converter against a dataset of 246 real-world COBOL programs, compiles and executes the generated Java, compares the output against expected results captured from the original COBOL programs, and reports detailed quality metrics.

The framework is not limited to COBOL-to-Java — any converter that implements the simple BaseCob2JavaConverter interface can be plugged in and evaluated.

---

How It Works

COBOL source
│
▼
Your Converter ──────────────────────────────────────────────────────┐
│ │
│ (optional: AST-enhanced) │
│ COBOL → AST extraction → LLM with structural context │
│ │
│ (base) │
│ COBOL → LLM → Java │
│ ▼
▼ Generated Java
javac (compile)
│
▼
java (run)
│
▼
Compare stdout & output files against expected ← captured from real GnuCOBOL run
│
▼
Metrics: exact match, cosine/Levenshtein/Jaccard similarity,
Halstead complexity, cyclomatic complexity, maintainability index, ...

Each test case is a test_metadata.json file that stores the COBOL source, input data files, stdin content, and the expected stdout/output files captured by running the original program through GnuCOBOL.

---

Results

Five converter/model combinations have been benchmarked on 244 test cases:

Converter	Model	Generation	Compilation	Run success	Cosine sim	Exact match
Base	Claude Sonnet 4.5	90.6%	90.6%	82.0%	0.638	17.3%
Base	Claude Haiku 4.5	73.4%	73.4%	65.2%	0.590	13.4%
AST	Claude Haiku 4.5	76.6%	76.6%	66.0%	0.545	11.2%
Base	Qwen 3 (local)	86.5%	86.5%	66.4%	0.472	8.2%
AST	Qwen 3 (local)	—	—	—	—	—

Generation = Compilation because the benchmark counts a test as compiled only when non-empty code was produced. The AST converter adds COBOL structural context to the prompt but does not consistently outperform the base approach — larger/smarter models benefit more from the extra context.

---

Prerequisites

• Python 3.10+
• Java JDK (for javac / java)
• GnuCOBOL (only needed to add new test cases, not to run the benchmark)
• An Anthropic API key or a GPU with enough VRAM for a HuggingFace model

---

Setup

1. Install Python dependencies

pip install -r requirements.txt

2. Configure environment variables

Create a .env file in the project root:

# Paths (defaults shown — override if needed)TESTS_DIR=testcasesCONVERSIONS_DEST_DIR=conversionsBENCHMARK_RESULTS_DIR=benchmark_resultsLOG_DIR=logs# Anthropic (if using Claude)ANTHROPIC_API_KEY=sk-ant-...ANTHROPIC_MODEL=claude-haiku-4-5# HuggingFace (if using a local model)MODEL_NAME=Qwen/Qwen2.5-Coder-7B-Instruct

3. (AST converter only) Set up Java libraries

bash setup.sh

Then manually download the Proleap COBOL Parser JAR from github.com/uwol/proleap-cobol-parser/releases and save it as lib/proleap-cobol-parser.jar. Finally compile the AST extractor:

javac -cp "lib/*" CobolAstToJson.java

---

Running the Benchmark

Open run_benchmark.py and edit the main() function — it is the single configuration point:

# 1. Describe the runrun_description="Base converter with Claude Sonnet 4.5"converter_description="BaseCobolConverter claude-sonnet-4-5"# 2. Choose your LLM backendfromutils.llm_anthropicimportLLM# Claude (API)# from utils.llm_huggingface import LLM # local HuggingFace model# 3. Choose your converterllm=LLM()
converter=BaseCobolConverter(llm) # simple one-shot conversion# converter = ASTCobolConverter(llm) # AST-enhanced conversion

Then run:

python run_benchmark.py

Results are written to benchmark_results/benchmark_results_<timestamp>.json. Intermediate per-test results are saved to conversions/conversions_<timestamp>/ after every test so the run is resumable.

Resuming an interrupted run

resume_previous_run=Trueprevious_conversions_dir=Path("conversions/conversions_20251122_140135")

---

Using Your Own Converter

Any converter that inherits from BaseCob2JavaConverter will work:

fromconverters.baseConverterimportBaseCob2JavaConverterclassMyConverter(BaseCob2JavaConverter):
defconvert(self, cobol_code: str) ->str:
# call your tool / API / rule-based engine herejava_code=my_translation_engine(cobol_code)
returnjava_code

Then pass it to the benchmark:

converter=MyConverter()
benchmark=Cob2JavaBenchmark(testcases_dir, conversions_dir, converter)
results=benchmark.run()

The LLM classes (utils/llm_anthropic.py, utils/llm_huggingface.py) are just helpers for the built-in converters — you do not need them if you bring your own converter.

---

Metrics

Success flags

Metric	Description
generation_success	Converter returned non-empty Java code
compilation_success	javac compiled the code without errors
run_success	Program ran and exited with code 0
conversion_time	Seconds taken to convert one program

Output similarity (expected vs. actual stdout/files)

Metric	Range	Interpretation
similarity_exact	0 / 1	Byte-exact match
similarity_cosine	0–1	Token-frequency similarity
similarity_levenshtein_similarity	0–1	Normalized edit distance (1 = identical)
similarity_levenshtein_distance	≥ 0	Raw edit distance (lower = better)
similarity_jaccard	0–1	Shared vocabulary ratio

Use similarity metrics together — high cosine + high Levenshtein means near-identical output; high cosine + low Jaccard suggests similar frequencies but different vocabulary.

Static code quality (generated Java)

Metric	Description
static_halstead_volume	Implementation size (operator/operand counts)
static_halstead_difficulty	Estimated cognitive difficulty
static_halstead_effort	Estimated implementation effort
static_cyclomatic_complexity	Number of independent paths (lower = simpler)
static_lines_of_code	Lines in the generated Java file
static_maintainability_index	Composite score 0–100 (higher = more maintainable)
static_maintainability_rating	Human-readable label (Good / Moderate / Poor)

All numeric metrics are reported as mean, min, max, and std across all test cases. Success flags are reported as sum (count) and ratio.

See metrics_explanation.md for detailed interpretation guidance.

---

Adding Test Cases

1. Place compilable .cbl files in src_data/ (verify they compile with GnuCOBOL first).
2. Filter out very short files:

   python utils/filter_short_files.py src_data short_files 60

3. Rewrite any absolute file paths in the COBOL source to be relative (e.g. data/experiment2/data.dat → data.dat).
4. Generate sample I/O metadata:

   python generate_sample_io_files.py

---

Project Structure

converters/
baseConverter.py ← abstract interface for all converters
cobolToJavaTranslator/
baseCobolConverter.py ← simple LLM converter
astCobolConverter.py ← AST-enhanced LLM converter
prompts.py ← conversion prompts
cobolAstParser.py ← COBOL → AST (via Java/ANTLR)
utils/
llm_anthropic.py ← Anthropic Claude wrapper
llm_huggingface.py ← HuggingFace Transformers wrapper
run_benchmark.py ← main entry point
testcases/success/ ← 246 test cases (test_metadata.json each)
benchmark_results/ ← final JSON reports per run
conversions/ ← per-test intermediate outputs