BIABench
收藏资源简介:
BIABench 是一个用于评估 AI agent 在真实世界端到端生物图像分析任务上的基准数据集。该数据集包含 16 个任务,这些任务均基于已发表的生物学研究重建,每个任务为 agent 提供原始图像和一份书面指令,并根据研究的真实标注和专家编写的评分标准对 agent 的输出进行评分。数据集的整体结构为:每个任务包含一个压缩的输入文件(input.zip)、一个评估文件(evaluation.zip,除一个任务外)、任务规范 YAML 文件(task_spec.yaml)、评估规则 YAML 文件(evaluation_rubric.yaml)以及包含文件校验和信息、来源和许可的 manifest.json 文件。下载总大小约为 11.1 GB(压缩),解压后约为 26.6 GB。16 个任务涵盖了多种生物图像模态,包括 H&E 组织学、荧光显微镜、相差显微镜、TIRF、共聚焦显微镜、光片显微镜等,维度从 2D 到 5D(时间序列、多通道、3D 等)。每个任务都从源数据集中选取了少量的图像(通常为 1 到 265 张),并进行了文件重命名或格式转换。该数据集主要用于评估和基准测试 AI agent 在生物图像分析中的能力,其地面真值数据仅用于评估目的,禁止用于模型训练或微调。数据集采用混合许可协议,不同任务可能有不同的图像许可(如 CC BY 4.0、CC BY-SA 4.0、CC0 1.0、GPL-3.0 等),但数据集本身的贡献(任务规范、评分规则、清单等)采用 CC BY 4.0 许可。
BIABench is a benchmark dataset for evaluating AI agents on real-world end-to-end biological image analysis tasks. The dataset contains 16 tasks, each reconstructed from published biological research, providing the agent with raw images and a written instruction, and scoring the agents output based on ground truth annotations and expert-written evaluation rubrics. The overall structure of the dataset: each task includes a compressed input file (input.zip), an evaluation file (evaluation.zip, except for one task), a task specification YAML file (task_spec.yaml), an evaluation rubric YAML file (evaluation_rubric.yaml), and a manifest.json containing file checksums, sources, and licenses. The total download size is approximately 11.1 GB (compressed) and about 26.6 GB when decompressed. The 16 tasks cover a variety of biological image modalities, including H&E histology, fluorescence microscopy, phase contrast microscopy, TIRF, confocal microscopy, light sheet microscopy, etc., with dimensions ranging from 2D to 5D (time series, multi-channel, 3D, etc.). Each task selects a small number of images from the source dataset (typically 1 to 265), with file renaming or format conversion. This dataset is primarily used for evaluating and benchmarking AI agents in biological image analysis; the ground truth data is for evaluation purposes only and is prohibited for model training or fine-tuning. The dataset adopts a mixed license: different tasks may have different image licenses (e.g., CC BY 4.0, CC BY-SA 4.0, CC0 1.0, GPL-3.0, etc.), but the contributions of the dataset itself (task specifications, scoring rules, manifests, etc.) are licensed under CC BY 4.0.
BIABench 数据集概述
基本信息
- 数据集名称:BIABench
- 数据集地址:https://huggingface.co/datasets/BIABench/BIABench
- 许可证:CC BY 4.0、CC BY-SA 4.0、CC BY 3.0、CC0 1.0、GPL-3.0、other(EMBL-EBI Terms of Use)
- 标签:biology、microscopy、bioimage-analysis、agents、benchmark、llm-agents
- 规模类别:n<1K
- 语言:en
- 查看器:false
数据集简介
BIABench 用于评估 AI 智能体在真实世界端到端生物图像分析任务中的表现。它包含 16 个任务,这些任务重建自已发表的生物学研究。每个任务为智能体提供原始图像和书面指令,并根据研究的真实标注和专家编写的评分标准对其输出进行评分。
- 网站:https://biabench.github.io
- 代码:https://github.com/BIABench/BIABench
- 论文:待添加
结构
<task_id>/
input.zip what the agent sees; unpacks to input/
evaluation.zip ground truth; unpacks to evaluation/ and carries evaluation/CANARY.txt
task_spec.yaml agent-facing specification (instructions, deliverables, source)
evaluation_rubric.yaml metric configuration and checklist filter used by the evaluator
manifest.json SHA-256 and size of every file, plus source and license
manifest.json aggregate of the per-task manifests, plus the canary string
LICENSE per-task license terms, sources, conditions and attribution wording
LICENSES/ full license texts referenced by the tasks
cytoplasm-nucleus-translocation-bbbc014 是唯一没有 evaluation.zip 的任务,因为它不附带参考标注。该任务改为根据智能体自身的每孔表格进行评分,通过测试源实验所确立的剂量反应来评分;evaluation_rubric.yaml 保存该配置。
下载大小为 11.1 GB 的压缩包,解压后为 26.6 GB。基准仓库的 benchmark_tasks/download_from_hf.py(https://github.com/BIABench/BIABench/blob/main/benchmark_tasks/download_from_hf.py)会验证 SHA-256 清单,并将每个任务解压到运行器和评估器所期望的 benchmark_tasks/<task_id>/{input,evaluation}/ 布局中。
任务
| task_id | short name | modality / dims | license of the images | input (unpacked) |
|---|---|---|---|---|
| he-nuinsseg-nuclear-segmentation | H&E nuclear segmentation | H&E histology, 2D | CC BY 4.0 | 4.0 MB |
| fluo-helacytonuc-cell-segmentation | HeLa nucleus/cytoplasm segmentation | fluorescence, 2D | CC BY 4.0 | 288 MB |
| fluo-cell-counting-2d-cellfmcount | IF cell counting | fluorescence, 2D | CC BY-SA 4.0 | 40 MB |
| smlm-localization-dnapaint | DNA-PAINT SMLM localization | TIRF, 2D time series | GPL-3.0 repository | 12.5 MB |
| cytoplasm-nucleus-translocation-bbbc014 | NF-kB translocation quantification | fluorescence, 2D | CC BY 3.0 Unported | 395 MB |
| microglia-phenotype-progression-bbbc054 | Microglia activation dynamics | phase contrast, 2D time-lapse | CC BY 3.0 Unported | 166 MB |
| fluo-coronavirus-golgi-colocalization | SARS-CoV-2 / Golgi colocalization | confocal, 2D | CC BY 4.0 | 247 MB |
| sim-microtubules-segmentation | Synthetic microtubule segmentation | simulated fluorescence, 2D | CC BY 4.0 | 7.2 MB |
| wound-healing-speed-kymograph-gigadb100118 | Wound-healing kymograph | DIC, 2D time-lapse | CC0 1.0 | 570 MB |
| fluo-dna-repair-foci-colocalization-sbsst227 | DNA repair foci colocalization | spinning-disk confocal, 2D time-lapse | EMBL-EBI Terms of Use (no CC license stated) | 733 MB |
| 3d-light-sheet-brain-vessels | LSFM brain capillary segmentation | light-sheet, 3D | CC BY 4.0 | 1.50 GB |
| confocal-mosaic-4channel-stitching-ctc-model | Multi-channel mosaic stitching | confocal, 4-channel 8x8 mosaic, 3 z | CC BY 4.0 | 1.62 GB |
| 5D-npc-assembly-kinetics-idr0115 | Nuclear pore reassembly dynamics | confocal, 3D time-lapse, 2 ch | CC BY 4.0 | 2.19 GB |
| 3d-confocal-puncta-quantification-sbiad1556 | Oncogenic puncta 3D quantification | spinning-disk confocal, 3D, 2 ch | CC0 1.0 | 5.09 GB |
| phase-contrast-bacteria-tracking-toiam | Bacteria single-cell tracking | phase contrast, 2D, 800 frames | CC BY 4.0 | 482 MB |
| 3d-fluo-cell-segmentation-lateral-line-idr0079 | Zebrafish lateral-line 3D cell segmentation | AiryScan confocal, 3D | CC BY 4.0 | 588 MB |
每个任务都是其源数据集的一个子集(通常为 1 到 265 张图像),并进行文件重命名,在少数情况下还会进行格式转换,例如将多帧影片存储为单帧 TIF。确切的选择以及每个文件的 SHA-256 都记录在 manifest.json 中;每个任务的源仓库和源论文记录在 LICENSE(https://huggingface.co/datasets/BIABench/BIABench/blob/main/LICENSE)以及任务自身的 YAML 中。
引用
bibtex @article{biabench2026, title = {BIABench: Evaluating AI agents on real-world bioimage analysis tasks}, author = {Pan, Zixuan and Panzeri, Davide and Johanns, Lukas and Moor, Marilin and Zhou, Yu and Peterson, Hedi and Shi, Yiyu and Chen, Jianxu}, year = {2026}, note = {Preprint. Venue, volume and DOI to be added.} }
请同时引用您所使用的每个任务的源数据集;若干许可证要求这样做。
许可证与署名
自有贡献(本卡片、任务规范、评分标准、清单以及派生真实标注)采用 CC BY 4.0;图像和上游真实标注保留上表所示的源许可证。各许可证要求的来源、条件和署名措辞记录在 LICENSE(https://huggingface.co/datasets/BIABench/BIABench/blob/main/LICENSE)中。
仅限评估通知
真实标注公开是为了让任何人都能运行评估器并复现论文的分数。每个 evaluation.zip 和聚合的 manifest.json 都包含金丝雀字符串 BIOIMAGEBENCH-CANARY-<uuid4>。请不要在 evaluation.zip 的内容上训练或微调模型,不要重新托管它们,并在任何派生副本中保持金丝雀字符串完整,以便检测污染。




