README.md

metadata

license: apache-2.0
tags:
  - benchmark
  - cuda
  - gpu
  - code-generation
  - orbench
size_categories:
  - 1B<n<10B

ORBench Test Data

Test data for ORBench — a benchmark for evaluating LLMs on CPU-to-CUDA code acceleration.

This repo hosts only the input data + expected outputs. The benchmark harness (CPU references, task_io adapters, evaluation scripts) lives at github.com/YOURNAME/ORBench.

Contents

File	Size	Description
small.tar.gz	~100 MB	Small inputs — for smoke tests (43 tasks)
medium.tar.gz	~3.5 GB	Medium inputs — main leaderboard
large.tar.gz	~8 GB	Large inputs — stress test
manifest.json	—	SHA256 + task metadata

Each tarball extracts to <task>/data/<size>/ containing:

• input.bin — binary input tensors (ORBench v2 format)
• expected_output.txt — reference output from CPU baseline
• cpu_time_ms.txt — CPU baseline wall time
• requests.txt — per-call queries (if applicable)

Usage

# Clone the harness repo
git clone https://github.com/YOURNAME/ORBench.git
cd ORBench

# Install HF client
pip install huggingface_hub

# Download data
python3 scripts/download_data.py small       # smoke
python3 scripts/download_data.py medium      # main leaderboard
python3 scripts/download_data.py large       # stress test
python3 scripts/download_data.py all         # everything

# Run the benchmark
python3 -m framework.run_all_tasks \
    --models gemini-3.1-pro-preview-openrouter \
    --levels 3 --sizes medium --yes

Verification

python3 scripts/download_data.py medium --verify

Citation

If you use ORBench, please cite:

@misc{orbench2026,
  title={ORBench: Evaluating LLMs on CPU-to-CUDA Code Acceleration},
  author={...},
  year={2026},
  url={https://github.com/YOURNAME/ORBench}
}