Expert Curated Data Samples

Explore high-quality datasets across key domains, purpose-built for real-world model development. Choose from ready-to-use packs or share a request for custom data solutions.

Explore datasets

Access Datasets

Latest Additions

Physical Intelligence

Structured, real-world datasets for robotic perception, action understanding, and environment-awaredecision-making.

Labeled taxonomy of eighteen rubric-writing skills across five categories and three proficiency levels for competency mapping and annotator calibration.

Multi-image reference to Video generation

Multi-reference images and edit prompts paired with human-scored 10-second videos assessed across six dimensions for video-generation evaluation.

Dataset Library

Explore curated data collections designed to accelerate model development across training, alignment, and evaluation workflows.

Vision Annotation

Images, video, and LiDAR point clouds annotated across nine types including bounding boxes, segmentation, and 3D cuboids for perception models.

View dataset ↗

Motion capture video samples

Seventy synchronized motion-capture samples pairing skeleton overlays, 3D renders, and egocentric video across household tasks for pose estimation.

View dataset ↗

Egocentric Hand Keypoint Annotation

First-person household video with 21-point 3D hand keypoints across cooking, cleaning, and other daily tasks for hand-pose estimation.

View dataset ↗

Egocentric video data

Twenty narrated first-person recordings of household and industrial tasks with think-aloud commentary for embodied action understanding.

View dataset ↗

Fifteen synchronized head- and wrist-camera recordings of manipulation tasks like clipping and threading for bimanual policy training.

View dataset ↗

Egocentric household annotated

Fifteen first-person household recordings with fine-grained hand and grasp annotations across domestic tasks for action recognition training.

View dataset ↗

Egocentric factory

Sixteen first-person clips of electronics assembly and PCB testing on an active production line for industrial task recognition.

View dataset ↗

Egocentric Factory De-IDed

Face-blurred first-person production-line video showing PCB handling and record-keeping for de-identified industrial action recognition.

View dataset ↗

Experts - Coding

Preference Pair Trajectory

Passing and failing coding-agent trajectories from matched Terminal Bench 2 tasks, paired for preference optimization and debugging-agent training.

View dataset ↗

Terminal bench tasks

Ten executable terminal tasks spanning workflows including data-pipeline debugging, systems programming, and security analysis, with deterministic verifiers for evaluation.

View dataset ↗

Cultural Reasoning

Native-expert prompts requiring local cultural knowledge, scored across 11 dimensions to evaluate multilingual model reasoning.

View dataset ↗

Language Experts

Indic-language prompts with four ranked model responses and rater explanations across six dimensions for preference-model training.

View dataset ↗

Language Experts

SFT Image + Text

Document, chart, receipt, table, and diagram images with ranked model responses and golden answers for vision-language model fine-tuning.

View dataset ↗

Language Experts

Linguistic Reasoning

Native-expert prompts testing grammar, morphology, syntax, semantics, and usage, scored across 11 dimensions for multilingual reasoning evaluation.

View dataset ↗

Voice SFT Samples

Scripted and unscripted multilingual speech recordings with verbatim transcripts for training voice-enabled speech and language AI systems.

View dataset ↗

Acoustic (Text-to-Audio) SFT

Indic scripted and spontaneous speech recordings across varied environments with word-for-word transcripts for speech generation and recognition models.

View dataset ↗

Multi-Turn Instruction Following (Long Context)

Seven multi-turn evaluations spanning 4 to 17 turns with failure signals localized across five dimensions including retrieval and reasoning.

View dataset ↗

Deep Research Agent Study - PCMB

Physics, chemistry, math, and biology prompts testing whether research agents propose novel next steps beyond existing literature.

View dataset ↗

LongHorizon STEM Expert Tasks

Expert-steered multi-turn STEM problems with pre-registered expected directions for evaluating model collaboration under scope changes and injected constraints.

View dataset ↗

Visual Designers

GUI grounding based code generation

Broken-and-fixed app screenshots across finance, health, media, board management, and travel paired with golden React code for UI repair.

View dataset ↗

Input-output image pairs with editing prompts spanning multiple categories for training image-to-image editing models.

View dataset ↗

Image-gen Thinking traces

Reasoning traces paired with prompts and resulting images, human-validated for training image models to reason before generating.

View dataset ↗

Multi-image reference to Video generation

Multi-reference images and prompts paired with 10-second generated videos scored across six dimensions for video-generation evaluation.

View dataset ↗

Spatial Counterfactual occlusion and rearrangement evaluation (SCORE)

Six spatial-editing workflows, including transparency, reflection, and object swapping, RLVR-scored for evaluating model spatial reasoning accuracy.

View dataset ↗

Video-to-video SFT Dataset

50,000 source-target video pairs with editing instructions spanning object changes, backgrounds, style, and effects for video-editing models.

View dataset ↗

Visual-Captioning

AI-generated captions paired with human corrections across video categories including industrial work, sports, and tutorials for caption-accuracy benchmarking.

View dataset ↗

Acoustic (Text-to-Audio) SFT

Indic scripted and spontaneous speech recordings across varied environments with word-for-word transcripts for speech generation and recognition models.

View dataset ↗

Labeled taxonomy of eighteen rubric-writing skills across five categories and three proficiency levels for competency mapping and annotator calibration.

View dataset ↗

Multi-image reference to Video generation

Multi-reference images and edit prompts paired with human-scored 10-second videos assessed across six dimensions for video-generation evaluation.

View dataset ↗

The Deccan AI Standard for Datasets

Explore datasets designed to be challenging, trustworthy, legally usable, and ready to scale.

Rigorous, Multi-Step Quality Control

Every sample passes through a purpose-built QC pipeline: human-led or hybrid approach using custom LLM judges with expert human validation.

Model-Breaking by Design

Datasets are benchmarked against five frontier models and accepted only when they expose meaningful failures in at least three.

Clean, Verifiable Usage Rights

Privately acquired data is sourced with clearly documented provenance and usage rights.

Built to Scale

Every creation pipeline is engineered to expand with your volume, complexity, and evolving model requirements.

Frequently Asked Questions

Sample datasets are typically delivered within 48 hours, and even sooner for off-the-shelf (OTS) data.

All sample data is shared through our data platform and can be downloaded, making it easy for your team to review, analyze, and assess fit for your workflow.

Absolutely. We can set up an expert consultation to understand your requirements and create a custom data pack tailored to your needs

Yes. We support both standard and highly specialized requests across domains, and can build data pipelines suited to your scope, format, and quality requirements.

This doesn’t have to end here

Accuracy is Intelligence

Access Datasets