Request evaluation access / 申请评测访问
Read the evaluation-use commitments and privacy notice, then describe your intended evaluation. / 请阅读评测用途承诺与隐私说明，并描述计划开展的评测。
Access is for evaluation. Do not use the requested controlled evaluation assets for pretraining, fine-tuning, distillation or reinforcement-learning training, and do not publicly redistribute them. Normal model evaluation is allowed. These commitments do not relicense source code or tools. See access-request.md and privacy-notice.md. / 访问用于评测：不将本次申请的受控评测资产用于预训练、微调、蒸馏或强化学习训练，不公开再分发；允许正常模型评测。这些承诺不重新许可源码或工具。
Log in or Sign Up to review the conditions and access this dataset content.
Website · Dataset · Task catalog · Citation · 中文
AccelerateBench
AccelerateBench provides 32 application-level program optimization tasks for evaluating coding agents on scientific and systems workloads. Each task includes application source, reference implementations, public checks and private-verifier implementations. Native Linux/Docker tools and a separate Nexus configuration layer support reproducible execution.
Request access
Describe your institution and intended evaluation in the access form. Independent researchers may enter “Independent.” Normal model evaluation, reproduction and auditing are permitted uses. Do not use the requested controlled evaluation assets for pretraining, fine-tuning, distillation or reinforcement-learning training, or publicly redistribute them.
Read Access request and commitments and Privacy notice. These access commitments are separate from software licenses: original source code and tools remain under their applicable licenses. Preserve the release's contamination notices and canary markers when managing evaluation copies.
Download
After approval, authenticate with hf auth login. Use Python 3.11 or later and an immutable dataset commit for reproducible downloads:
python3 -m venv .venv
. .venv/bin/activate
python -m pip install 'huggingface_hub==1.8.0'
export HF_DATASET_ID='AdAstraAbyssoque/accelerate-bench'
export HF_REVISION='COMMIT_SHA'
hf download "$HF_DATASET_ID" accelerate-bench-public32-nexus-maintenance-canary-v2.tar.gz SHA256SUMS --repo-type dataset --revision "$HF_REVISION" --local-dir ./acceleratebench-download
cd acceleratebench-download
sha256sum -c SHA256SUMS # macOS: shasum -a 256 -c SHA256SUMS
mkdir accelerate-bench
tar -xzf accelerate-bench-public32-nexus-maintenance-canary-v2.tar.gz -C accelerate-bench
cd accelerate-bench
python -m pip install -r requirements-tools.txt
python tools/generate.py --task datalife-io-monitoring --model example-model
Replace COMMIT_SHA with the commit from the dataset's revision history. The final command generates a configuration offline. For native execution, follow public-release/INSTALL-NATIVE.md; for Nexus configuration, read adapters/nexus/README.md in the extracted package.
The complete source asset is 2,291,316,141 bytes with SHA-256 d91457a7597b9830abdf3440a5388eed14cf160a57dccbf3d38c4e9c01554aa8. Source-cohort metadata binds the package to its canonical source and release marker.
Layout
| Repository file | Contents |
|---|---|
accelerate-bench-public32-nexus-maintenance-canary-v2.tar.gz |
Complete task, source and environment package |
task-catalog.json |
Task list, resource profiles and license/hydration pointers |
SHA256SUMS, source-cohort.json |
Download integrity and version binding |
canary-metadata.json |
Release marker and contamination-notice binding |
LICENSE, notices/ |
Component licenses and preserved notices |
| Directory inside the package | Contents |
|---|---|
canonical/tasks/ |
32 tasks: original application source, checks, scoring and resource budgets |
profiles/native/, public-release/ |
Native preparation tools and installation guidance |
profiles/nexus/, adapters/nexus/ |
Separate environment profiles and Nexus input generation |
tools/ |
Offline configuration generation, capacity planning and intent tracking |
canaries.json, CONTAMINATION_NOTICE.md, contamination/ |
Release and per-task canary markers and notices |
Two ICON grids use the documented official hydration procedure. Preserve each task's cases, scoring, 8-hour agent budget and verifier resources. Report native and Cloud compatibility cohorts separately: Nexus outer CPU budgets are shared and are not equivalent to nested native hard limits.
Licenses
Original author-owned source code and tools use Apache-2.0. The manuscript uses CC BY 4.0. Bundled third-party code, data and models retain their component licenses, including ABySS's noncommercial GPL-derived terms, ICON file-level exceptions and HPL acknowledgements. The aggregate is therefore labeled license: other.
See LICENSE, licensing scope, third-party inventory and scope resolutions. Access approval does not grant additional third-party rights or change software licenses.
Citation
Use CITATION.cff for the complete author list and manuscript metadata.
@unpublished{liu_acceleratebench,
title = {Accelerate Bench: Evaluating Agents for Application-Level Program Optimization},
author = {Liu, Bowen and Li, Yuhan and Chen, Nuo and Luo, Weile and Peng, Miao and Zhang, Qintong and Zhang, Qifan and Li, Jia and Gao, Ziqi and Huang, Lydell},
url = {https://acceleratebench.com/}
}
中文说明
AccelerateBench 提供 32 个应用级程序优化任务，涵盖科学计算与系统工作负载。每个任务包含应用源码、参考实现、公开检查与私有判题实现，并配有原生 Linux/Docker 工具和独立的 Nexus 配置层。
申请与下载
在访问申请表中填写单位、用途类别和具体评测用途；独立研究者可填写“Independent”。允许正常模型评测、复现与审计；本次申请的受控评测资产不用于预训练、微调、蒸馏或强化学习训练，不公开再分发。阅读访问承诺及隐私说明。这些承诺不重新许可源码和工具。
获准后用 hf auth login 登录，按上方 Download 命令下载固定 commit、校验 SHA256 并解包。包内 canonical/tasks/ 为 32 题原始源码和判题，profiles/ 与 adapters/ 为环境配置层，contamination/ 保存发布 canary 与逐题说明。ICON 两份网格按官方 hydration 流程获取。
保留原病例、评分、8 小时模型预算与判题资源；原生和 Cloud 兼容环境分组报告，Nexus 外层共享 CPU 预算不等于原生嵌套硬限额。自有代码和工具采用 Apache-2.0，论文采用 CC BY 4.0，第三方许可逐项保留。引用见上方 BibTeX 或 CITATION.cff。
- Downloads last month
- 40