datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Zebra-CoT
Zebra‑CoT
A diverse large-scale dataset for interleaved vision‑language reasoning traces.
Dataset Description
Zebra‑CoT is a diverse large‑scale dataset with 182,384 samples containing logically coherent interleaved text‑image reasoning traces across four major categories: scientific reasoning, 2D visual reasoning, 3D visual reasoning, and visual logic & strategic games.
Dataset Structure
Each example in Zebra‑CoT consists of:
Problem statement:… See the full description on the dataset page: https://huggingface.co/datasets/multimodal-reasoning-lab/Zebra-CoT.ZebraLogicPaper: https://huggingface.co/papers/2502.01100
Arxiv: https://arxiv.org/abs/2502.01100
Citation
@article{zebralogic2025,
title={ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning},
author={Bill Yuchen Lin and Ronan Le Bras and Kyle Richardson and Ashish Sabharwal and Radha Poovendran and Peter Clark and Yejin Choi},
year={2025},
url={https://arxiv.org/abs/2502.01100},
}
@article{dziri2024faith,
title={Faith and fate: Limits of transformers on… See the full description on the dataset page: https://huggingface.co/datasets/WildEval/ZebraLogic.YOLO-OBB-Zebra-Crossings-Datasettextlatent_zebra_thinkmorph_armAB
Text-Latent (Arm A) vs All-Latent (Arm B) — Zebra-CoT + ThinkMorph
35638 samples/arm, 18 categories. Schema = ULVR/williamium style (sample_id, category, source_dataset,
question, answer, input_image, intermediate_image_N, num_intermediate_steps, messages_json).
armA_text_latent: real decoded text CoT + latent visual blocks (intermediate_image_1..3).
armB_render_latent: reasoning text RENDERED to images, all-latent baseline (intermediate_image_1..17).
messages_json = full Monet… See the full description on the dataset page: https://huggingface.co/datasets/RuoliuYang/textlatent_zebra_thinkmorph_armAB.ZebraLogicBenchmulti-zebra-logic
Dataset Card for the MultiZebraLogic dataset
This dataset includes zebra puzzles in 39 European and 5 non-European languages and in two sizes: 2x3 and 4x5. It can be used for evaluating logical reasoning ability.
The data has been generated using the code in this repo.
Dataset Details
Dataset Description
Zebra puzzles are a type of constraint satisfaction problem. They describe a number of objects, N_objects, that each have attributes… See the full description on the dataset page: https://huggingface.co/datasets/alexandrainst/multi-zebra-logic.spatial-mmcot-zebra_multihop
Spatial MMCoT v1 · zebra_multihop
Zebra-CoT multi-hop object counting over rendered 3D scenes (primitive objects on textured ground under a sky). Many traces open with a viewpoint change, so the first target is often a novel view. The final thought re-examines the last target image, but on some rows it only says it counts the objects and never states the number; the count is then only in <answer>. Released rows carry 2 to 5 target images. 2,057 traces with more target images… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_multihop.spatial-mmcot-zebra_tetris
Spatial MMCoT v1 · zebra_tetris
Zebra-CoT Tetris: polyomino puzzles of three kinds: apply a sequence of transformations to a shape, tile a shape with a set of pieces, and fill the grid outside a shape. The options are drawn in the input image, so the answer is a letter. Each step has its own target image: transformation puzzles start with a redraw of the start shape, then one image per transformation; tiling puzzles start with the isolated shape, fill-the-complement puzzles with… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_tetris.ZebraLogicBench-privatespatial-mmcot-zebra_jigsaw
Spatial MMCoT v1 · zebra_jigsaw
Zebra-CoT visual jigsaw, type-2 (single-image) rows only, on ImageNet images (mostly photographs; some are web graphics such as banners and flyer templates). Each row has one input image: the source image with its missing piece(s) greyed out, above a panel of four candidate piece sets labelled A-D. The options exist only in that image, so the answer is the letter. The target is the complete source image. Type-1 rows are excluded: three quarters of… See the full description on the dataset page: https://huggingface.co/datasets/yrlyrl/spatial-mmcot-zebra_jigsaw.Zebra-CoT-unify-stylezebra-cot-mistral-small-3.2-24b-preprocessed
Zebra-CoT Preprocessed — Mistral Hackathon 2026
Preprocessed version of the Zebra-CoT dataset for fine-tuning Mistral-Small-3.2-24B-Instruct.
Format
text: formatted as [INST] question [/INST] <think> reasoning </think> answer
image: PIL JPEG image for the corresponding visual task
Usage
Fine-tuning Mistral-Small-3.2-24B on chain-of-thought visual reasoning.
Hackathon
Created for Mistral Hackaton 2026 — Fine-tuning track with W&B.
zebra-puzzlesSynthetic data for the paper [2505.05755] Insertion Language Models: Sequence Generation with Arbitrary-Position Insertions.
Project page: https://dhruveshp.com/projects/ilm
rg-zebra_puzzles-instruct-100k
RLVR generated dataset
Procedural rows from reasoning-gym, formatted for verl GRPO.
Build metadata
{
"config": "/home/owais/Projects/rlvr/rlvr/configs/datasets/zebra_puzzles-instruct.yaml",
"template_type": "qwen-instruct",
"developer_prompt": null,
"data_source": "reasoning_gym",
"default_extract": "answer_tag",
"train_rows": 100000,
"test_rows": 4096,
"train_seed": 42,
"test_seed": 43,
"tasks": {
"zebra_puzzles": {
"weight": 1… See the full description on the dataset page: https://huggingface.co/datasets/carbonteq/rg-zebra_puzzles-instruct-100k.zebra-puzzles-sortedzebra-kb-explanations
ZEBRA: Zero-Shot Example-Based Retrieval Augmentation for Commonsense Question Answering
A retrieval augmentation framework for zero-shot commonsense question answering with LLMs.
🛠️ Installation
Installation from PyPi
pip install zebra-qa
Installation from source
git clone https://github.com/sapienzanlp/zebra.git
cd zebra
conda create -n zebra python==3.10
conda activate zebra
pip install -e .
🚀 Quick Start… See the full description on the dataset page: https://huggingface.co/datasets/sapienzanlp/zebra-kb-explanations.ZebraArena
ZebraArena
Dataset accompanying the paper ZebraArena: A Diagnostic Simulation Environment for Studying Reasoning–Action Coupling in Tool-Augmented LLMs.
ZebraArena is a procedurally generated diagnostic environment for studying reasoning–action coupling in tool-augmented LLMs, with controllable difficulty and a knowledge-minimal design. Each task is a partially observed Zebra (logic-grid) puzzle: a Constraint Satisfaction Problem with a unique ground-truth solution, where a subset… See the full description on the dataset page: https://huggingface.co/datasets/ZebraArena/ZebraArena.math_logic_zebralogic_trainlm-eval-results-mlabonne-Zebrafish-7B-private
Dataset Card for Evaluation run of mlabonne/Zebrafish-7B
Dataset automatically created during the evaluation run of model mlabonne/Zebrafish-7B
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-mlabonne-Zebrafish-7B-private.logic__zebra_puzzle_dataset_200Puzzle_Zebra_AllPuzzle_Zebra_20Krlvr_logic__zebra_puzzle_1.3kZebraArena
ZebraArena
ZebraArena dataset.
Configs missing1 … missing7 correspond to the number of missing clues; splits small / medium / large correspond to puzzle size.
Usage
from datasets import load_dataset
ds = load_dataset("WanjiaAlicia/ZebraArena", "missing3", split="large")
Puzzle_Zebra_20K_completionszebra_100
Zebra 100
Overview
This synthetic micro-dataset contains 177 pure logic deduction traces (5x5 zebra puzzles), consisting of 100 training examples and 77 validation examples. It features strictly logic grid puzzles and includes zero mathematical data.
The 100 training examples were randomly sampled without replacement from the train split of tamewild/instruct5, while the 77 validation examples are kept exactly the same across both datasets.
It was created to… See the full description on the dataset page: https://huggingface.co/datasets/tamewild/zebra_100.math_logic_puzzles_zebralogic_level_4Puzzle_Zebramath_logic_puzzles_zebralogic_level_5math_logic_puzzles_zebralogic_level_1