Rockman Benchmark v0.2

A comprehensive, contamination-resistant benchmark for evaluating LLMs on code generation, debugging, optimization, and software engineering tasks.

---

🎯 Overview

Rockman is a production-ready benchmark designed to rigorously evaluate Large Language Models on realistic software engineering tasks. Unlike simple code generation benchmarks, Rockman covers the full spectrum of developer workflows:

Task Type	Description	Examples
Generation	Implement function from signature + docstring	"Implement Dijkstra's algorithm"
Debugging	Fix bugs in provided code	Off-by-one, edge cases, infinite loops
Optimization	Meet specific complexity targets	O(n log n), O(1) space
Stateful	Classes with persistent state	LRU cache, banking ledger, event processor
Multi-step	Parse → Build → Solve pipelines	Parse graph → build adjacency → find path

---

📊 Benchmark Composition (v0.2)

Split	Tasks	Hidden Tests	Purpose
Public	154	❌	Development, debugging
Private	44	✅ (encrypted)	Official evaluation
Hidden	22	✅ (never published)	Final leaderboard integrity

Category	Tasks	Difficulty Range
Fundamentals	60	1–3
Data Structures	50	2–4
Algorithms	50	2–5
Graphs	60	3–6
Trees	40	3–5
Math	40	2–5
Debugging	50	2–4
Complexity	30	3–5
Software Engineering	40	3–5
Stateful Systems	30	4–6
Multi-step	30	4–6
Total	520	1–7

Difficulty Levels: 1=Introductory → 7=Research

---

🔐 Contamination Prevention

Rockman implements six layers of contamination protection:

Layer	Mechanism
1. Canary Strings	Unique ROCKMAN_CANARY_<hash> per task
2. Content Hashing	SHA256 of prompt + tests for integrity
3. Split Isolation	Public/Private/Hidden with zero overlap
4. Encrypted Hidden Tests	AES-256-GCM, keys held by evaluator only
5. Source Attribution	source_type, source_url, derived_from metadata
6. Temporal Controls	Creation/publication dates, embargo periods

Verification: Run rockman verify --dataset rockman_v0.2.jsonl to audit integrity.

---

🚀 Quick Start

Installation

git clone https://github.com/InserLoft/Rockman
cd Rockman
pip install -e .
pip install matplotlib  # for PNG reports

Run Baseline Evaluation

# Phase 1: Reference baselines (always works offline)
python phase1_evaluate_baselines.py

# Phase 2: Model evaluation via OpenRouter
python phase2_evaluate_openrouter.py

Programmatic Usage

from rockman import (
    Evaluator, Problem, load_problems_from_jsonl,
    RockmanResultsReporter, generate_benchmark_dataset
)

# Load dataset
problems = load_problems_from_jsonl("rockman_v0.2_private.jsonl")

# Evaluate model
evaluator = Evaluator(benchmark_version="v0.2")
completions = {p.task_id: [your_model(p.prompt)] for p in problems}
results = evaluator.evaluate_dataset(problems, completions, k=1)

# Generate reports
reporter = RockmanResultsReporter("results")
reporter.generate_full_report(leaderboard, detailed_results)

---

📈 Evaluation Protocol

Parameter	Value	Description
Temperature	0.0	Deterministic generation
Top-p	1.0	No nucleus sampling
Max Tokens	4096	Sufficient for complex tasks
k (Pass@k)	1	Single sample per task
Timeout	2.0s	Per-task execution limit
Memory	256 MB	Per-task memory limit
Sandbox	Containerized	Docker (or local fallback)

Required Metadata per Evaluation

{
  "model_name": "user/model",
  "model_version": "model-version",
  "provider": "provider",
  "temperature": 0.0,
  "top_p": 1.0,
  "max_tokens": 4096,
  "k": 1,
  "hardware": "example-cloud",
  "date": "2024-09-20T01:48:31Z",
  "evaluator_version": "0.2.0",
  "benchmark_version": "v0.2"
}

API Error Handling

Critical: Network/API errors are not counted as code failures.

Error Type	Recorded As	Counted as Failure
Code assertion failure	FAIL	✅ Yes
Runtime error	FAIL	✅ Yes
Timeout	TIMEOUT	✅ Yes
API 429 (rate limit)	RATE_LIMITED	❌ No
API 402/400 (payment)	API_ERROR	❌ No
API 404/500	API_ERROR	❌ No

---

📊 Official Reports

Every evaluation automatically generates:

Format	Description
results/*.csv	Complete breakdown: overall, difficulty, category, task type, language, per-problem
results/*.png	Professional charts (black bg, white primary): leaderboard, difficulty bars, category bars, task type donut

Sample Leaderboard Output

ROCKMAN BENCHMARK v0.2
Private Evaluation — 44 Tasks

Model                    Pass@1    Solved
------------------------------------------------
Template                 13.64%     6/44
Heuristic                 0.00%     0/44
Random                    0.00%     0/44

---

🏗️ Architecture

rockman/
├── benchmark/           # Core evaluation engine
│   ├── schema.py        # Problem, Score, Result dataclasses
│   ├── evaluator.py     # Unified evaluator (all task types)
│   ├── metrics.py       # Pass@k, Wilson CI, statistical tests
│   ├── runners/         # Language runners (Py, C++, Java, JS, TS, Rust, Go)
│   ├── sandbox.py       # Docker sandbox + local fallback
│   ├── encryption.py    # AES-256-GCM hidden test encryption
│   ├── versioning.py    # Immutable version registry
│   └── determinism.py   # Reproducibility tracking
├── dataset/             # Dataset generation & management
│   ├── generator.py     # Template-based task generation
│   ├── categories.py    # 12 categories, 50+ subcategories
│   ├── validators.py    # 13 validation checks (incl. negative testing)
│   └── splits.py        # Public/Private/Hidden splits + canaries
├── tasks/               # Specialized task types
│   ├── debugging.py     # Bug injection (8 bug types)
│   ├── complexity.py    # Performance test generation
│   ├── stateful.py      # LRU, banking, event processor, etc.
│   └── multi_step.py    # Parse→Build→Solve pipelines
├── baselines/           # Reference baselines
│   ├── calibration.py   # Difficulty calibration
│   ├── human_baseline.py # Human evaluation framework
│   └── reference_models.py # Random, Heuristic, Template
├── reports/             # Report generation
│   └── generate.py      # Markdown/HTML/JSON reports
├── visualization/       # CSV + PNG report generation
│   └── results_reporter.py # Black-theme professional charts
├── cli/                 # CLI commands (generate, eval, leaderboard, publish)
├── docs/                # Methodology, scoring, security, contamination
│   ├── methodology.md
│   ├── scoring.md
│   ├── security.md
│   └── contamination.md
├── phase1_evaluate_baselines.py
├── phase2_evaluate_openrouter.py
├── freeze_v0.2.py
├── rockman_benchmark.py   # Backward-compatible facade
├── rockman_cli.py
└── rockman_v0.2_manifest.json

---

🔧 CLI Commands

# Generate dataset
rockman generate --count 500 --encrypt-hidden --output rockman_v0.2.jsonl

# Evaluate model
rockman evaluate --dataset rockman_v0.2_private.jsonl --completions-file model_outputs.json --k 1

# Generate leaderboard
rockman leaderboard --results-file eval_results.json

# Publish dataset
rockman publish --dataset rockman_v0.2.jsonl --public-output public.jsonl --private-output private.jsonl

# Validate dataset
rockman validate --dataset rockman_v0.2.jsonl --solutions-file refs.json

---

📦 Dataset Access

Platform	Link	Access
Hugging Face	Inserloft/Rockman	datasets.load_dataset("Inserloft/Rockman")
GitHub	InserLoft/Rockman	git clone + releases

from datasets import load_dataset
ds = load_dataset("Inserloft/Rockman", split="private")
# Note: Private split requires authentication + master secret for hidden tests

---

📜 License

MIT License — see LICENSE for details.

Dataset splits:

• Public: CC-BY-4.0
• Private/Hidden: Restricted (evaluation only)

---

🏷️ Citation

@misc{rockman2024,
  title={Rockman: A Comprehensive Benchmark for Code Generation and Software Engineering},
  author={Inserloft},
  year={2026},
  version={v0.2},
  url={https://github.com/InserLoft/Rockman}
}

---

🤝 Contributing

1. Fork the repository
2. Create a feature branch
3. Add tasks via rockman generate or PR new task templates
4. Ensure all 13 validation checks pass
5. Submit PR with validation report

---

📞 Contact

Platform	Link
Website	Inserloft.com
Benchmarks	Inserloft.com/benchmarks/RockMan
GitHub	InserLoft/Rockman
Hugging Face	Inserloft/Rockman

---

Inserloft — Advancing the science of code intelligence evaluation.