DATASETS · MULTIMODAL & BENCHMARK-TARGETED
Off-the-shelf, high-quality datasets for Generative AI, Physical AI, and World Models — collected in the real world, verified and annotated by domain experts, delivery-ready for large-model training.
OTS · 01
Starting from gameplay video, captures player–environment interaction data via low-level reverse engineering of games, producing "UI-free, highly synchronized, ground-truth-action" world-model training data.
OTS · 02
Collects problems requiring multi-step reasoning (causal judgment, logic puzzles); experts decompose them into minimal reasoning units, build reasoning-flow diagrams, design adversarial samples to catch logical traps, and produce multiple solution versions.
OTS · 03
Covers coding-task generation + multi-turn dialogue + evaluation, plus the Vibe Coding project's standardized end-to-end workflow for rating LLM coding capability.
OTS · 04
Text-to-image/video foundation-model post-training, plus a multimodal RL aesthetics-ranking project (1 prompt + 7 generated images, ranked across multiple dimensions per group).
OTS · 05
Covers mobile, desktop, and web operation screenshots spanning social, e-commerce, office, finance, education, and a dozen more real task categories, each step annotated with the reasoning, the action taken, and the target UI element — built specifically for agent and computer-use model training.
OTS · 06
200,000 high-definition anime creative videos, backed by a full AIGC production pipeline from script and image generation through video generation and editing, and a dedicated team delivering 30–50 episodes/month, scalable to 50–80 soon.
Accelerate your model’s capabilities with our suite of pre-curated, off-the-shelf datasets. Purpose-built to plug directly into your training pipelines, these high-fidelity assets are explicitly engineered to drive immediate performance lifts across core reasoning, coding, and domain-specific benchmarks.
DATASET OVERVIEW
Designed to evaluate agents' long-horizon task execution in real-world professional workflows, ALE provides expert-level evaluation data with economic relevance and verifiable outcomes. It supports agent training and evaluation, as well as research into cross-industry workflow deployment, long-horizon planning, and verifiable execution.
DATA PRODUCTION STRENGTHS
We have expert-level data production capabilities for long-horizon task scenarios. Through ongoing collaboration with the ALE team and data contribution, we have access to multi-domain task design and review resources supported by industry experts and a distributed contributor network. Our workflow covers task design, sandbox environment construction, hidden references, deterministic scorer development, and quality review.
DATASET OVERVIEW
Designed for long-horizon software engineering task solving by frontier coding agents, DeepSWE provides original, contamination-free engineering task data. The dataset covers five programming languages, with tasks created from scratch and independently verified based on software behavior, enabling evaluation of real-world engineering capabilities rather than memorization. It can be used for coding agent training, capability evaluation, and failure-mode research.
DATA PRODUCTION STRENGTHS
We have dual-source data acquisition capabilities covering open-source repositories and authorized private repositories, together with expertise in designing original long-horizon tasks across languages and repositories. Our capabilities cover sandboxed trajectory collection, behavior-level test development, and patch-level independent verification, supporting the full workflow from task creation and execution to quality control and scoring.
DATASET OVERVIEW
Designed to evaluate coding agents on large-scale, cross-file code refactoring, SWE-bench Pro Max provides expert-reviewed, multilingual software engineering evaluation data. It is used to assess the real-world upper bound of AI coding models on complex, long-horizon software engineering tasks.
DATA PRODUCTION STRENGTHS
We have expert-level data production capabilities for large-scale, cross-file code refactoring scenarios. Through close collaboration with the SWE-bench Pro Max team, we have access to professional task design and review resources across multiple programming languages. Our workflow covers task design, repository-instance selection, issue-description rewriting, test-suite review, reference-patch alignment, and behavioral preservation verification, supported by an automated data-production pipeline for scalable production and expert-based quality evaluation.
DATASET OVERVIEW
Designed to evaluate agents' end-to-end ability to perform professional computing tasks in isolated terminal environments, the dataset provides challenging and continuously extensible evaluation tasks. It supports terminal agent training and evaluation, as well as research on tool use, environment interaction, and failure modes.
DATA PRODUCTION STRENGTHS
We have expert-level, cross-disciplinary task production capabilities for challenging terminal-based scenarios, with professional task designers across software engineering, scientific computing, finance, engineering, and system operations. Our workflow covers task design, containerized environment construction, reference solution development, terminal-state verifier development, instruction–verification alignment, Oracle solvability validation, and reward-hacking resistance review, supporting scalable production and quality assurance.
DATASET OVERVIEW
Designed to evaluate frontier models on expert-level interdisciplinary knowledge and closed-form reasoning, HLE provides original, verifiable, and retrieval-resistant academic question-answering data. It covers mathematics, natural sciences, humanities, and social sciences, and is used to evaluate deep reasoning, knowledge mastery, and confidence calibration.
DATA PRODUCTION STRENGTHS
We have expert-level, interdisciplinary capabilities for producing challenging academic problems, supported by domain experts and professional task design and review resources. Our capabilities cover task design, answer uniqueness verification, retrieval-resistance validation, and quality review, supporting scalable production and automated evaluation of high-difficulty domain-specific data.
DATASET OVERVIEW
Designed for challenging real-world engineering tasks, the dataset provides specification-driven system implementation data, covering design specifications, engineering implementation, and correctness verification. It can be used for coding agent training, evaluation, and research.
DATA PRODUCTION STRENGTHS
We have expert data production capabilities across software development, operations, DevOps, and system design, supported by domain-specific engineering experts and review resources. Lead domain experts ensure technical correctness, while independent reviewers and evaluation specialists conduct secondary verification. Our workflow covers complete project construction, reproducible environment setup, task and reference implementation development, public/hidden test design, and rubric development.
