Request evaluation access / 申请评测访问

Read the access commitments and privacy notice on the dataset card, then describe your intended evaluation. / 请先阅读数据集页面上的访问承诺与隐私说明,再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. The full access commitments and privacy notice are on the dataset card. / 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。完整的访问承诺与隐私说明见数据集页面。

Log in or Sign Up to review the conditions and access this dataset content.

Accelerate Bench

Website Leaderboard arXiv GitHub Harbor Hub

Accelerate Bench asks whether coding agents can make whole scientific applications faster without breaking them. This dataset holds the 32 tasks of v1.1, the version evaluated in the paper, as Harbor task packages.

Commit-derived tasks compared with Accelerate Bench's workload-defined tasks

Commit-derived tasks (left, as in GSO) start from one known performance commit and measure the agent's patch against it. Accelerate Bench tasks (right) fix the workload and the computation that must be preserved; the agent decides where and how to optimize, and a separate verifier scores the final program.

中文说明

Each task gives a coding agent a working scientific application, a fixed workload and a CPU and memory budget. The agent has eight hours to make the whole program faster while every required output stays correct; a separate private verifier then rebuilds the final program, checks it on hidden inputs and times it against the unmodified program.

  • Tasks: 32, from 8 application areas
  • Format: Harbor task directories with task.toml, instruction.md, environment/ and tests/
  • Version: v1.1, the version evaluated in the paper
  • Download: accelerate-bench-v1.1-harbor-tasks.tar.gz, 2.56 GB
  • Harbor Hub: 31 of the tasks (all except ICON) are public as accelerate-bench/accelerate-bench
  • Metric: Score@1, correctness-gated runtime savings averaged over the 32 tasks
  • Code: downloader, verifier and runner on GitHub

Access

To run the tasks without a request, use Harbor Hub; it has every task except ICON.

The full archive in this repository has all 32 tasks and a manifest of every file's SHA-256. Request access with the form above: give your institution (or “Independent”) and your planned evaluation. Approval is manual. Model evaluation, reproduction, auditing and teaching are supported. Please do not use the tasks for pretraining, fine-tuning, distillation or reinforcement-learning training, and do not redistribute the archive publicly.

Access commitments

Access requests support benchmark evaluation, reproduction, auditing and teaching; normal evaluation of models is allowed. Please provide your institution or organization (“Independent” if unaffiliated), a purpose category and a short description of the intended evaluation, without confidential project details or sensitive personal information. Hugging Face provides the account information used for the request.

For the controlled evaluation assets requested through this repository, you confirm that you will not use them for model pretraining, fine-tuning, distillation or reinforcement-learning training, and that you will not publicly redistribute them.

These are access-use commitments for controlled evaluation assets. They do not relicense source code or tools, remove rights granted by existing component licenses, or add a commercial-use restriction. Original author-owned code and tools remain Apache-2.0; third-party components keep their licenses. The manuscript's CC BY 4.0 license is separate.

Privacy notice

Requests are reviewed by Bowen Liu, the maintainer of this repository (bliu690@connect.hkust-gz.edu.cn). Applications are managed in Hugging Face; the account information and form answers are used only to review requests, manage access and contact applicants about them. They are not published in the repository, the dataset, benchmark outputs, the paper, the leaderboard or public logs, and no offline bulk exports are made.

The form asks only for an institution, a purpose category, a short evaluation description and two confirmations. Do not provide passwords, tokens, identity documents, billing details or confidential project information. Contact the maintainer to correct or delete any maintainer-held copies; Hugging Face's own records follow its policies.

Run from Harbor Hub

Harbor downloads the 31 public tasks itself. It needs Linux x86-64 with Docker:

pip install harbor==0.22.0
export ANTHROPIC_API_KEY=...
harbor run -d accelerate-bench/accelerate-bench@v1.1 \
  -a claude-code -m anthropic/MODEL_ID --ak reasoning_effort=high -n 1

-n 1 runs one task at a time. Each task gives the agent 4 or 8 CPUs and up to 32 GiB of memory and runs a public-test service next to it, so raise -n only if the host has room for several tasks at once. Add -i hpcg-multigrid-pcg to run a single task. ICON is not on Harbor Hub because its two grid files have to be downloaded from their official source.

Download the full archive

After approval, the recommended route is the downloader in the code repository, which checks the archive and every task file before installing them into canonical/tasks/:

git clone https://github.com/AdAstraAbyssoque/accelerate-bench
cd accelerate-bench
pip install -r requirements-download.txt
hf auth login
python scripts/download_data.py

To download the archive directly instead:

hf download AdAstraAbyssoque/accelerate-bench \
  accelerate-bench-v1.1-harbor-tasks.tar.gz SHA256SUMS \
  --repo-type dataset --revision 9b7825dbee991573edbc290f6effaffbffff8d7e \
  --local-dir ./accelerate-bench-data
cd accelerate-bench-data
sha256sum -c SHA256SUMS        # macOS: shasum -a 256 -c SHA256SUMS
tar -xzf accelerate-bench-v1.1-harbor-tasks.tar.gz

ICON's two grid files are downloaded from their official source; the quickstart has the command and shows how to run a task with Harbor.

Leaderboard

Benchmark v1.1, one run per agent and task, with all final programs scored on the same machine (paper Table 1). Score@1 and Aτ are percentages.

# Agent Harness Score@1 GM A1.25 A2
1 Claude Sonnet 5.5 Claude Code 67.4 3.83× 100.0 71.9
2 Claude Opus 5.5† Claude Code 59.4 3.94× 84.4 68.8
3 GPT-6.1 Sol Codex 49.2 2.42× 78.1 50.0
4 GPT-6 Astra Codex 47.2 2.40× 75.0 53.1
5 Hy4-Preview Claude Code 45.2 2.31× 71.9 50.0
6 Hy4 Claude Code 45.0 2.41× 75.0 46.9
7 Kimi K3 Claude Code 42.5 2.47× 68.8 40.6
8 DeepSeek-V4.1-Flash-1017 Claude Code 40.1 2.28× 65.6 34.4
9 GLM-5.3 Claude Code 39.7 2.20× 65.6 40.6
10 GLM-5.3-Flash Claude Code 39.3 2.20× 68.8 34.4
11 DeepSeek-V4-Pro Claude Code 37.9 2.19× 62.5 34.4
12 GPT-6.1 Terra Codex 37.9 1.92× 75.0 25.0
13 DeepSeek-V4-Flash Claude Code 34.1 1.92× 56.2 34.4
14 GLM-5.2 Claude Code 34.0 1.98× 46.9 37.5
15 Grok 4.6 Claude Code 33.9 2.13× 50.0 34.4
16 Hy3 Claude Code 25.5 1.76× 37.5 28.1

†The provider's safety system stopped the Claude Opus 5.5 runs on Graph sparsification, PRESTO and VPIC; they count as failures with zero reward. On its other 29 tasks, Claude Opus 5.5 scores 65.6%. Gemini-3.8-Flash is not listed because its v1.1 runs are not yet complete.

The task reward is r = C × max(0, 1 − Tcandidate / Tbaseline), where C is 1 if the program passes the verifier's correctness checks. Score@1 is the mean reward over the 32 tasks; Aτ is the share of tasks that are correct and at least τ× faster; GM is the geometric mean speedup, counting incorrect or unaccelerated outcomes as 1×.

Tasks

Task Directory Area CPUs Memory Verifier limit
GROMACS gromacs-molecular-dynamics Molecules & materials 4 6 GiB 390 min
LAMMPS lammps-molecular-dynamics Molecules & materials 4 8 GiB 30 min
GPAW gpaw-fd-dft-restart Molecules & materials 4 6 GiB 233 min
Quantum ESPRESSO qe-scf-nscf-dos Molecules & materials 4 6 GiB 233 min
SIESTA siesta-dft-restart Molecules & materials 4 6 GiB 233 min
Fluidity fluidity-heat-checkpoint Fluids & atmosphere 4 8 GiB 233 min
ICON icon-atmosphere-checkpoint Fluids & atmosphere 4 6 GiB 200 min
MPAS-A mpas-atmosphere-restart Fluids & atmosphere 4 6 GiB 233 min
NEMO nemo-lock-exchange Fluids & atmosphere 4 8 GiB 233 min
OpenFOAM openfoam-cavity-piso Fluids & atmosphere 4 8 GiB 40 min
Palabos palabos-lattice-boltzmann Fluids & atmosphere 4 6 GiB 120 min
SU2 su2-transonic-euler Fluids & atmosphere 4 8 GiB 233 min
OpenMC openmc-multigroup-transport Transport & astronomy 4 6 GiB 120 min
ShengBTE shengbte-phonon-transport Transport & astronomy 4 6 GiB 233 min
VPIC vpic-harris-reconnection Transport & astronomy 4 6 GiB 150 min
PRESTO presto-pulsar-search Transport & astronomy 4 6 GiB 120 min
ABySS abyss-genome-assembly Bioinformatics 4 8 GiB 30 min
MHM2 mhm2-metagenome-assembly Bioinformatics 4 6 GiB 150 min
MrBayes mrbayes-phylogenetics Bioinformatics 4 6 GiB 120 min
wtdbg2 wtdbg2-long-read-assembly Bioinformatics 4 8 GiB 30 min
HPCG-derived hpcg-multigrid-pcg Numerics & graphs 4 6 GiB 120 min
HPL-derived hpl-dense-lu Numerics & graphs 4 6 GiB 120 min
Graph sparsification fegrass-graph-sparsification Numerics & graphs 4 4 GiB 120 min
miniVite minivite-louvain Numerics & graphs 4 6 GiB 120 min
NUFFT imaging nufft-imaging Numerics & graphs 4 6 GiB 200 min
Tucker reconstruction tucker-reconstruction Numerics & graphs 8 16 GiB 120 min
QuEST quest-circuit-simulation Quantum circuits 4 8 GiB 60 min
QiboTN qibotn-tensor-network-circuits Quantum circuits 8 32 GiB 833 min
Paged LM generation paged-lm-generation Machine learning 4 6 GiB 200 min
ThunderSVM thundersvm-train-predict Machine learning 4 8 GiB 60 min
DataLife datalife-io-monitoring Scientific I/O 4 6 GiB 233 min
IO500-derived io500-file-io Scientific I/O 4 6 GiB 120 min

Every task gives the agent eight hours. CPUs and memory are the agent's allocation; the verifier limit is the private verifier's time limit. HPCG-derived, HPL-derived and IO500-derived are adapted tasks, not official competition benchmark scores.

What is in this repository

File Contents
accelerate-bench-v1.1-harbor-tasks.tar.gz The 32 task packages under canonical/tasks/, canonical/manifest.json (path, size, mode and SHA-256 of every task file), licenses, citation metadata and contamination markers
third-party-sources/fluidity-ppa-sources.tar Source packages of the Ubuntu packages bundled with the Fluidity task
SHA256SUMS, third-party-sources/SHA256SUMS Checksums of the two archives
task-catalog.json, source-cohort.json, canary-metadata.json, canaries.json Task list, release provenance and contamination markers
LICENSE, notices/ Component license terms and third-party notices

Version 1.1 starts from task source 20261010-v5-timing3 and includes the later v1.1 revisions of MHM2 (task version 1.1.1: statement, correctness checks and reference source), Fluidity (output parser; Ubuntu dependencies bundled in .fluidity-ppa/), SU2 (verifier time limit) and IO500 (system-call trace parser). The v1.1 release notes describe each change.

Licenses

Original author-owned code and tools use Apache-2.0; the manuscript uses CC BY 4.0. Bundled third-party code, data and models keep their component licenses, including ABySS's noncommercial GPL-derived terms, ICON file-level exceptions and HPL acknowledgements; the aggregate is therefore labeled license: other. Several Ubuntu packages bundled with the Fluidity task are LGPL or GPL licensed, and their source packages are in third-party-sources/. Access approval does not grant additional third-party rights.

Contamination markers

The release carries one release-level and 32 task-level canary markers, including the one at the top of this card. Please keep them when managing evaluation copies, and exclude the tasks from training corpora. The code repository includes a scanner (tools/scan_canaries.py); a hit shows that the text is present, not that a model was trained on it.

Citation

@misc{liu2026acceleratebench,
  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},
  year   = {2026},
  eprint = {2611.20882},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url    = {https://arxiv.org/abs/2611.20882}
}

中文说明

Accelerate Bench 评测编码 agent 能否在保持结果正确的前提下,让完整的科学计算应用跑得更快。本数据集为论文评测所用的 v1.1 版本,共 32 道题,以 Harbor 任务包的形式发布(约 2.56 GB)。每道题给出可运行的应用、固定工作负载和 CPU/内存预算,agent 有 8 小时优化;私有验证器重新构建最终程序,在隐藏输入上检查正确性,并与未修改的程序交替计时。

Harbor Hub: 除 ICON 外的 31 道题已公开发布为 Harbor 数据集 accelerate-bench/accelerate-bench,无需申请,按上方 harbor run -d accelerate-bench/accelerate-bench@v1.1 命令即可运行。ICON 的两份网格文件需从官方来源下载,因此只在完整发布包中提供。

申请访问: 完整发布包含全部 32 道题及每个文件的 SHA-256 清单。在上方表单填写单位(无隶属可填 Independent)与计划开展的评测,人工审核。允许模型评测、复现、审计与教学;请勿将题目用于预训练、微调、蒸馏或强化学习训练,不公开再分发发布包。完整的访问承诺与隐私说明见上方折叠内容。

下载与运行: 获批后推荐使用代码仓库中的 scripts/download_data.py 下载并逐文件校验;也可按上方命令用 hf download 下载固定版本并核对 SHA256。运行方法见 GitHub 快速上手。

版本与许可: v1.1 在任务源码 20261010-v5-timing3 的基础上,包含 MHM2、Fluidity、SU2 和 IO500 四题的 v1.1 修订。自有代码采用 Apache-2.0,论文采用 CC BY 4.0,第三方组件保留原许可;Fluidity 附带的 Ubuntu 软件包中有 LGPL/GPL 组件,对应源码包见 third-party-sources/。

Downloads last month
227

Paper for AdAstraAbyssoque/accelerate-bench