datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
wholebrain_crispr_atlas
PerturbAI Brain-Wide In Vivo CRISPR Atlas
This dataset represents a landmark in functional genomics: spanning 8 million single cells in living tissue and hundreds of distinct neuronal cell types, this is the most expansive in vivo functional genomics resource ever created. By mapping the language of biology at an unprecedented scale, our platform provides the foundation for the next generation of AI-driven therapeutic discovery.
Manuscript: “Genome-scale functional mapping of… See the full description on the dataset page: https://huggingface.co/datasets/perturbai/wholebrain_crispr_atlas.chem-perturbridge
Chem-PerturBridge
Processed standardized expression .h5ad files and differential gene expression (DGE) .h5ad outputs from Chem-PerturBridge datasets whose upstream terms allow redistribution.
This repository contains:
processed group-replicate bulk or pseudobulk input files under <dataset>/
DGE outputs under <dataset>/group_rep/ and <dataset>/sep_rep/
The DGE replicate modes indicate whether the DGE analysis was run using all replicates or per replicate (for within-dataset… See the full description on the dataset page: https://huggingface.co/datasets/theislab/chem-perturbridge.PerturbDiff_dataefficientnet-v2-l-adv-dataset
Perturb Adversarial Images
Verified adversarial examples for efficientnet_v2_l (torchvision/EfficientNet_V2_L_Weights.IMAGENET1K_V1), produced by the
Perturb network. Each row is one clean image together with all of its
verified adversarial versions: images that are imperceptibly different from the original
(L∞ ≤ 0.03 in [0,1] pixel scale) yet change the model's top-1 prediction.
This dataset grows continuously. New rows are appended as the network produces them and uploaded in… See the full description on the dataset page: https://huggingface.co/datasets/perturb-ai/efficientnet-v2-l-adv-dataset.Perturb-Sapiens
Perturb Sapiens: A Human Whole-Organism Atlas of Perturbed Cells
Dataset Description
Perturb Sapiens is an evolving database of AI-predicted single-cell perturbation responses, representing the first human whole-organism atlas of perturbed cells.
Perturb Sapiens is generated using the post-trained Stack model (Stack-Large-Aligned), an in-context learning foundation model for single-cell biology.
Data Sources:
Prompt Data: Parse/OpenProblems PBMC perturbation data
Query… See the full description on the dataset page: https://huggingface.co/datasets/arcinstitute/Perturb-Sapiens.perturbseq_normalized
CellClip public normalized perturbation single-cell release
This public repository contains the source-cleared, human, cell-level portion
of CellClip Stage 1: 254 H5AD files, 12,718,270 cells, and
260,859,318,437 payload bytes. These are processed derivatives rather than
the original raw download archives.
Source partition
Units
Cells
Reference cells
Effect cells
scPerturb (23 genepert + 229 chempert)
252
4,774,802
397,528
4,377,274
XCell / X-Atlas Orion (genepert)… See the full description on the dataset page: https://huggingface.co/datasets/cfy2yue/perturbseq_normalized.MultiBBQ-perturbations
MultiBBQ: image perturbations
Image-level perturbation sets used for the robustness experiments in Fairness Failure
Modes of Multimodal LLMs. Each set is the GPT-Image-1 image collection from
MLL-Lab/MultiBBQ with a single, controlled
transform applied. Evaluating on a perturbed set measures how stable a model's fairness
behavior is under everyday image degradations.
Paper: Fairness Failure Modes of Multimodal LLMs
Code:… See the full description on the dataset page: https://huggingface.co/datasets/MLL-Lab/MultiBBQ-perturbations.perturb_ubuntu_osworld_file_cacheperturbenchThe dataset contains data used in work:
"Perturbench: Benchmarking machine learning models for cellular perturbation analysis."
The data comes from the following publications:
Norman, T. M., Horlbeck, M. A., Replogle, J. M., Ge, A. Y., Xu, A., Jost, M., Gilbert, L. A., and Weissman, J. S. (2019). Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science, 365(6455):786–793.
Srivatsan, S. R., McFaline-Figueroa, J. L., Ramani, V., Saunders, L., Cao, J., Packer… See the full description on the dataset page: https://huggingface.co/datasets/altoslabs/perturbench.PerturbMultiTwo datasets, namely diet_conditions_20240411 and crispr_screen_20240615, are included. Image files for each
dataset are archived into several .tar files. Inside each .tar file, the name of each image file is
the ID of the cell.
The imaging channels of each file are Alb, polyT, rRNA, M6PR, CathB, Perilipin,
Sqstm1, LC3b, TOMM20, Calreticulin, pS6RP, Na+/K+ATPase, SNAP23, TOM70, Rab7, mtRNA, Vimentin, Gapdh.
The RNA expression levels, spatial coordinates, and conditions of individual cells… See the full description on the dataset page: https://huggingface.co/datasets/xingjiepan/PerturbMulti.perturb-seq-pseudo-pairing-benchmarks
Perturb-seq Pseudo-pairing Benchmarks
Dataset Summary
This repository provides processed single-cell perturbation transcriptomic datasets and representative pseudo-control pairings used to study how pseudo-control construction affects perturbation modeling.
Single-cell perturbation assays are destructive: the same cell cannot be observed before and after perturbation. Cell-level modeling therefore requires an estimated or sampled unperturbed counterpart, referred… See the full description on the dataset page: https://huggingface.co/datasets/JFLa/perturb-seq-pseudo-pairing-benchmarks.whatsup_all_perturbedvcell-perturbation-source-data
ConvergeCELL Source Datasets — v1.0.0
Cached H5ADs of every single-cell RNA-seq dataset registered in the
ConvergeCELL data catalog. Mirrors the original sources (GEO, figshare,
Tabula Sapiens, CellxGene) so downstream code has a single, fast, versioned
endpoint to fetch from.
Each row of every h5ad is one cell; each column is one gene (HGNC symbol).
The exact obs/var schema follows whatever the original source provided —
this bundle does not re-annotate, harmonize, or QC. For a… See the full description on the dataset page: https://huggingface.co/datasets/nicolas-lynn/vcell-perturbation-source-data.foundation-models-perturbationData for the paper "Foundation Models Improve Perturbation Response Prediction" as described on GitHub.
vcc-perturb
Virtual Cell Challenge 2025 H1 hESC training-set perturbation atlas
CRISPRi (gene knockdown) in H1 human embryonic stem cells. Single-cell expression in log-normalized counts.
Generated 2026-06-08 as one of three companion atlases (Norman, Replogle, VCC).
File schema (each config / single-config repo)
File
Shape
Description
pseudobulks.h5ad
(50, n_genes)
50 control pseudobulks (15 cells each, log-normalized means). Cell-type-specific baseline.… See the full description on the dataset page: https://huggingface.co/datasets/nicolas-lynn/vcc-perturb.GUI-Perturbed
GUI-Perturbed
A step-level GUI grounding dataset built on domain-randomized web pages for diagnosing visual and spatial heuristics in VLM agents.
📄 Technical Report · 🌐 Baseline Result Viewer · 💻 Code
Overview
GUI-Perturbed is an evaluation dataset for step-level GUI element localization. It is designed to expose and precisely diagnose the failure modes of vision-language model (VLM) GUI agents to examine whether models rely on rigid visual shortcuts rather… See the full description on the dataset page: https://huggingface.co/datasets/figai/GUI-Perturbed.boolq-natural-perturbationsBoolQ questions with semantic alteration and human verifications
@article{khashabi2020naturalperturbations,
title={Natural Perturbation for Robust Question Answering},
author={D. Khashabi and T. Khot and A. Sabhwaral},
journal={arXiv preprint},
year={2020}
}
terminal_bench_2_perturbed_docker_exp_freelancer_tasks_glm_4_7_traces_20260223_182635norman-perturb
Norman 2019 CRISPRa K562 perturbation atlas
CRISPRa (gene activation) in K562 chronic myeloid leukemia cells. Single-cell expression in log-normalized counts.
Generated 2026-06-08 as one of three companion atlases (Norman, Replogle, VCC).
File schema (each config / single-config repo)
File
Shape
Description
pseudobulks.h5ad
(50, n_genes)
50 control pseudobulks (15 cells each, log-normalized means). Cell-type-specific baseline.
coexpression.h5ad… See the full description on the dataset page: https://huggingface.co/datasets/nicolas-lynn/norman-perturb.replogle-perturbperturbation-embeddings
Gene embeddings for perturbation modeling
This repository distributes three existing gene embedding tables in a common
NumPy/JSON format for use with pertTF and other perturbation models. These are
gene-level features, not measured perturbation responses or pertTF-generated
predictions. The original embedding models were not trained by this release.
Tables
Directory
Representation
Rows
Dimensions
Published dtype
esm2
ESM2 protein embeddings distributed… See the full description on the dataset page: https://huggingface.co/datasets/weililab/perturbation-embeddings.b1k_perturb_recovery3_task0_curatedtask417_mickey_es_sentence_perturbation_generation
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task417_mickey_es_sentence_perturbation_generation
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task417_mickey_es_sentence_perturbation_generation.perturbed_humanevalPerturbed version of HumanEval from: ReCode: Robustness Evaluation of Code Generation ModelsOpenR1-Math-220k-pruned-head-random-perturbationPerturbReason
🧬 AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling(ACL 2026 Findings)
📃 Paper • 🐙 Code • 🤗 Model
Please refer to our repository and paper for more details.
🌐 Overview
PerturbReason is the training data for the AROMA model. AROMA is a novel multimodal architecture for virtual cell modeling that integrates textual evidence, graph topology, and protein sequences to predict the effects of genetic… See the full description on the dataset page: https://huggingface.co/datasets/blazerye/PerturbReason.LEGIT-VIPER-Jigsaw-Toxic-Comment-Perturbedeval_PerturbA_OpenLoopThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "so100",
"total_episodes": 1,
"total_frames": 877,
"total_tasks": 1,
"total_videos": 3,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:1"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/shylee/eval_PerturbA_OpenLoop.AIME-2021-2025-Perturbed-Solutions-V1.0-Gemini-2.5-Flashparc2026-track2-perturb-dataset-20260923
track2 摂動カテゴリ学習データセット(PARC2026 Track2、2026-09-22〜23)
PARC2026 本選 Track2(LIBERO-plus の摂動つきタスク、π0.5 の微調整用)の教師データ。公開 8 タスクの元タスク × 公開例題に出る摂動カテゴリ 4 つ
(Sensor Noise / Camera Viewpoints / Robot Initial States / Objects Layout)を、P+D 制御の scripted expert の軌道で被覆したもの。
全 6,711 本、1,262,626 フレーム。すべて成功・非対象物の L1 変位 1 mm 以下・300 手以内・把持点と置く点の通り越し 1.1 mm 以下。
作り方・教師の設定・欠番の理由・途中で直した欠陥は manifests/TRACK2_PERTURB_DATASET_20260923.html(日本語)に書いてある。
中身
ディレクトリ
本数
フレーム
中身… See the full description on the dataset page: https://huggingface.co/datasets/ekunish/parc2026-track2-perturb-dataset-20260923.