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.OpenMath-Vision-CoT-10klibero_cotThis dataset was created using LeRobot.
It contains embodied Chain-of-Thought (CoT) demonstrations for the LIBERO benchmark, featuring paired reasoning and action traces. It was curated as part of the DeepThinkVLA project using a two-stage data engine that distills key frames with a cloud LVLM and scales to full trajectories via a fine-tuned local VLM.
Paper: DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models
Repository: https://github.com/OpenBMB/DeepThinkVLA… See the full description on the dataset page: https://huggingface.co/datasets/yinchenghust/libero_cot.CottonWeedDet12
Dataset Card for CottonWeedDet12
CottonWeedDet12 is a 12-class weed object-detection dataset for cotton production systems in the southern U.S., consisting of 5,648 RGB field images with 9,370 bounding box annotations collected in Michigan State University MEFAS field trials during 2021-2022. It is the companion dataset for the YOLOWeeds benchmark of YOLO object detectors.
This is a FiftyOne dataset with 5648 samples.
Installation
If you haven't already, install… See the full description on the dataset page: https://huggingface.co/datasets/Voxel51/CottonWeedDet12.Reason-RFT-CoT-Dataset
🤗 Reason-RFT CoT Dateset
The full dataset used in our project "Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning".
⭐️ Project │ 🌎 Github │ 🔥 Models │ 📑 ArXiv │ 💬 WeChat
🤖 RoboBrain: Aim to Explore ReasonRFT Paradigm to Enhance RoboBrain's Embodied Reasoning Capabilities.
♣️ Quick Start
Please refer to Dataset Preparation
🔥 Overview
Visual reasoning abilities play a crucial role in understanding complex multimodal… See the full description on the dataset page: https://huggingface.co/datasets/tanhuajie2001/Reason-RFT-CoT-Dataset.GenRef-wds
GenRef-1M
We provide 1M high-quality triplets of the form (flawed image, high-quality image, reflection) collected across
multiple domains using our scalable pipeline from [1]. We used this dataset to train our reflection tuning model.
To know the details of the dataset creation pipeline, please refer to Section 3.2 of [1].
Project Page: https://diffusion-cot.github.io/reflection2perfection
Dataset loading
We provide the dataset in the webdataset format for fast… See the full description on the dataset page: https://huggingface.co/datasets/diffusion-cot/GenRef-wds.glaucoma-expert-cot-final
Glaucoma Expert Chain-of-Thought
Ophthalmologist six-step reasoning reports for fundus photographs, each paired with
a binary glaucoma label. 1,074 cases from LAG and Papila.
Files
file
rows
glaucoma / not
train.jsonl
823
304 / 519
val.jsonl
92
46 / 46
test.jsonl
159
79 / 80
images/
1,074
<source>_<id>.jpg
Record schema
{
"id": "1689",
"source": "LAG",
"image": "LAG_1689.jpg",
"split": "train",
"final_diagnosis_GT":… See the full description on the dataset page: https://huggingface.co/datasets/yuzhench/glaucoma-expert-cot-final.cothanhaudiocloudflare_imgBed_publiclibero_cotThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "panda",
"total_episodes": 1693,
"total_frames": 273465,
"total_tasks": 40,
"total_videos": 0,
"total_chunks": 2,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:1693"
},
"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/Fitz135/libero_cot.HumanRef-CoT-45k
🦖🧠 Rex-Thinker: Grounded Object Referring via Chain-of-Thought Reasoning 🦖🧠
We propose Rex-Thinker, a Chain-of-Thought (CoT) reasoning model for object referring that addresses two key challenges: lack of interpretability and inability to reject unmatched expressions. Instead of directly predicting bounding boxes, Rex-Thinker reasons step-by-step over candidate objects to determine which, if any, match a given expression.… See the full description on the dataset page: https://huggingface.co/datasets/IDEA-Research/HumanRef-CoT-45k.cots_yolo_dataset
🪸 CSIRO Crown-of-Thorns Starfish (COTS) Detection Dataset — YOLO Format
This dataset is a modified version of the CSIRO COTS and COTS Scars Dataset, originally released under the Creative Commons Attribution 4.0 License (CC BY 4.0).
The original dataset contains images and annotations for Crown-of-Thorns Starfish (COTS) and COTS scars, collected to support coral reef monitoring and control efforts on the Great Barrier Reef (GBR).
These starfish are coral predators, and their… See the full description on the dataset page: https://huggingface.co/datasets/eloise54/cots_yolo_dataset.GenRef-CoT
GenRef-CoT
We provide 227K high-quality CoT reflections which were used to train our Qwen-based reflection generation model in ReflectionFlow [1]. To
know the details of the dataset creation pipeline, please refer to Section 3.2 of [1].
Dataset loading
We provide the dataset in the webdataset format for fast dataloading and streaming. We recommend downloading
the repository locally for faster I/O:
from huggingface_hub import snapshot_download
local_dir =… See the full description on the dataset page: https://huggingface.co/datasets/diffusion-cot/GenRef-CoT.zebra-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.
tbg-cot-bench
TBG-CoT-Bench
TBG-CoT-Bench is a local application benchmark for testing temporal belief tracking over Chain-of-Thought-style evidence sequences.
The benchmark evaluates whether a system can track belief about the temporal claim:
Event A occurred before Event B.
This repository contains synthetic temporal reasoning scenarios, rule-based baselines, local EXAONE/Ollama experiments, trajectory visualizations, generated reports, and pytest-based application benchmark checks.… See the full description on the dataset page: https://huggingface.co/datasets/CHML-real/tbg-cot-bench.vlmn_iphonecf100_cotrain_magicsoup_no_insta_sub5
vlmn_iphonecf100_cotrain_magicsoup_no_insta_sub5
Description
VLN Navigation dataset with 100% of counterfactual iphone data, 10% of magicsoup no insta subsampled to 5 points.
Processing Parameters
{}
Dataset Configuration
Train dataset:
mixer: mateoguaman/coda_every1_25pct_sub5: 0.1
mateoguaman/iphone_every1_sub5: 1.0
mateoguaman/scand_every1_50pct_sub5: 0.1
mateoguaman/spot_every1_sub5: 0.1
mateoguaman/tartandrive_every1_100pct_sub5: 0.1… See the full description on the dataset page: https://huggingface.co/datasets/mateoguaman/vlmn_iphonecf100_cotrain_magicsoup_no_insta_sub5.polaris_droid_cotrainThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "panda",
"total_episodes": 315,
"total_frames": 49528,
"total_tasks": 17,
"total_videos": 0,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 15,
"splits": {
"train": "0:315"
},
"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/imishutk0410/polaris_droid_cotrain.glaucoma-expert-cot-raw-1077
Glaucoma Expert Chain-of-Thought
Ophthalmologist six-step reasoning reports for fundus photographs, each paired with
a binary glaucoma label. 1,074 cases from LAG and Papila.
Files
file
rows
split
expert_cot_trainval.jsonl
915
train (823) + val (92)
expert_cot_test.jsonl
159
test
images/
1,074
<source>_<id>.jpg
Record schema
{
"id": "1689",
"source": "LAG",
"image": "LAG_1689.jpg",
"split": "train"… See the full description on the dataset page: https://huggingface.co/datasets/yuzhench/glaucoma-expert-cot-raw-1077.AIME25-CoT-CN
Sci-Bench-AIME25'
This repo is a branch of Sci Bench made by IPF team. Mainly include the AIME 25' solution with multi-modal CoT and diverse solving path.
Brief intro
💻 Overview
A brief template and final report will be posted in Isaac's Blog
And the markdown template can be found in data/I_2
❓ Why we do this?
The multi-lingual datasets are scarce, while the CoT of Math is even less, no matter whether the CoT or the solution contains pictures… See the full description on the dataset page: https://huggingface.co/datasets/IPF/AIME25-CoT-CN.Resume-Analysis-CoTR
Resume Reasoning and Feedback Dataset
Dataset Description
This dataset contains approximately 417 examples designed to facilitate research and development in automated resume analysis and feedback generation. Each data point consists of a user query regarding their resume, a simulated internal analysis (chain-of-thought) performed by an expert persona, and a final, user-facing feedback response derived solely from that analysis.
The dataset captures a two-step reasoning… See the full description on the dataset page: https://huggingface.co/datasets/Akhil-Theerthala/Resume-Analysis-CoTR.Zebra-CoT-unify-stylelibero_cot_openvlaLMOD-Cataract-1K-surgical-analysis-cot
Cataract-1K LLM-Generated Surgical Instructions
Dataset Overview
This dataset is derived from the Cataract-1K dataset (part of the LMOD benchmark) and enhanced using Qwen3-VL-30B-A3B-Thinking, a large vision-language model with reasoning capabilities. It is designed for training medical AI systems to provide actionable surgical guidance with transparent reasoning.
Generation Process
Source Data: Cataract-1K processed frames with segmentation annotations… See the full description on the dataset page: https://huggingface.co/datasets/mehti/LMOD-Cataract-1K-surgical-analysis-cot.libero_cot_contiousMeta-CoT-Training-Data-ExampleOpenSpatialDataset_CoTlibero_spatial_cotThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "panda",
"total_episodes": 1693,
"total_frames": 273465,
"total_tasks": 40,
"total_videos": 0,
"total_chunks": 2,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:1693"
},
"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/jayshim/libero_spatial_cot.RESA-CoT-data
RESA-CoT Dataset
The RESA-CoT dataset is a multimodal dataset designed for large language model alignment and reasoning research. It consists of image-conversation pairs in LLaVA format, enhanced with Chain-of-Thought (CoT) style reasoning to improve interpretability and alignment.
Dataset Versions
RESA
Based on VLGuard data.
Augmented using GPT-4o to generate CoT-style conversations.
RESA-mix
Combines RESA with 10K LLaVA-NEXT samples.
Also enhanced with CoT-style… See the full description on the dataset page: https://huggingface.co/datasets/yfwang22/RESA-CoT-data.R-CoTLLaVA-CoT-Instruct-58KThis dataset is borrowed from LLaVA-CoT-o1-Instruct