W32 AI Data Intelligence

One-line Summary

NVIDIA released multiple robot video datasets on October 9 [P0], Google DeepMind released the human-verified personalized multimodal benchmark Life-Bench [P0], and synthetic data and spatial world models gained momentum simultaneously in China's open-source community [P1]. The strongest data demand signal this week: robot task trajectory and failure review.

Key Findings

This week's 5 high commercial value findings

P0 NVIDIA Releases Multiple Robot Video Datasets on October 9 [P0]

On 2026-10-09, NVIDIA released or updated three Agile One robot video datasets: `nvidia/agile_one_s_pick_up_ssd_54eps` with 2 downloads and 0 likes, `nvidia/agile_one_s_pick_196_eps` with 349 downloads and 0 likes, and `nvidia/Agile_one_s_PNP_Dataset` with 84 downloads and 0 likes. During the same period, NVIDIA continued to hold highly popular Physical AI data assets: `nvidia/PhysicalAI-SmartSpaces` with 85,614 downloads and 93 likes, `nvidia/PhysicalAI-Robotics-NuRec` with 3,314 downloads and 81 likes, and `nvidia/PhysicalAI-Traffic-Anomaly-Reasoning` with 1,250 downloads and 20 likes.

Business significance → Robot and Physical AI data is shifting from “general-purpose scene video” toward “specific task episode trajectories,” such as pick, pick-and-place, and anomalous traffic reasoning. The value of this data lies not only in collection, but also in human judgment of whether an action succeeded, the reasons for failure, object-contact relationships, and the boundaries of anomalous events. Knowlyr can prioritize high-judgment-density data services such as robot task trajectory review, failure-case triage, and long-tail event attribution.

P0 Google DeepMind Releases Human-Verified Personalized Multimodal Benchmark Life-Bench [P0]

On 2026-10-07, Google DeepMind released `google/life-bench`, with 62 downloads and 3 likes. The dataset is described as a fully synthetic, human-verified multimodal benchmark, with a scale description beginning with 11,81, targeting multimodal personalization beyond concept recognition. Task labels include visual-question-answering, question-answering, and image-text-to-text, with image and text as the modalities. DeepMind's model `google/embeddinggemma-2` was released on 2026-09-14, with 29,185 downloads and 1,354 likes, and is positioned as an open lightweight multimodal embedding model.

Business significance → The key signal from Life-Bench is that “synthetic data still requires human verification,” especially in personalization, multimodal question answering, and everyday-life scene understanding, where correct answers often depend on context, preferences, and commonsense judgment. For Knowlyr, the opportunity is not low-level labeling, but building a “human judgment validation layer”: assessing whether models truly understand user intent, personal preferences, image-text contextual consistency, and whether their answers align with real-world decision-making logic.

P1 Synthetic Data and Spatial World Models Gain Momentum Simultaneously in China's Open-Source Community [P1]

On 2026-10-09, InternLM released `internlm/SpatialASM-edit140k`, with 0 downloads and 0 likes. Its tags include image-text-to-image, Spatial Understanding, Multi-image, World Model, and Novel-view Generator. On the same day, it released `internlm/SpatialASM-128k`, also with 0 downloads and 0 likes. On the Chinese open-source foundation model side, Qwen released `Qwen/Qwen-Image-2.1-Turbo` on 2026-10-09, with 303 likes and a text-to-image pipeline. The synthetic category included four datasets: `nvidia/Open-SWE-Traces`, `internlm/SpatialASM-128k`, `laion/Vivarium-Greenscreen-Characters-FLUX-9B`, and `laion/Anim-E-Ghibli-Backgrounds-FLUX-9B`.

Business significance → Chinese model teams focused this week on image generation, spatial understanding, and world model data, with an emphasis on generation and application deployment. The core bottleneck for synthetic spatial data is quality evaluation: whether viewpoints are physically consistent, whether editing preserves object identity, and whether spatial relationships are reasonable. Knowlyr can productize “spatial consistency judgment,” “multi-image editing preference ranking,” and “world model error diagnosis” to serve image generation and embodied AI teams.

P1 Coding Agent Data Shifts from Trajectory Collection toward Capability Decomposition and Contamination Control [P1]

NVIDIA's `nvidia/Open-SWE-Traces` has 25,904 downloads and 146 likes. Version 1.2 was released on 2026-09-26, adding agent trajectories generated by Qwen3.8-27B for mini-swe-agent and previewing trajectories from OpenCode and Claude Code harnesses. The paper `Code Understanding is a Bottleneck for Coding Agents` was released on 2026-10-07 and introduced CABRA: Coding Ability Blueprint for Rigorous Agent evaluation, which constructs tasks using call graph transformations. In a concurrent blog post, iMerit’s `When Coding Benchmarks Stop Being Reliable and What to Build Instead` highlighted answer leakage and contamination issues in coding benchmarks. On GitHub, `anthropics/claude-code` has 149,889 stars, `openai/codex` has 128,402 stars, `openai/skills` has 27,963 stars, and `anthropics/skills` has 180,178 stars.

Business significance → Coding Agents have entered a phase where “SWE-bench scores alone are not enough.” The market needs verifiable, reproducible, contamination-resistant task construction and trajectory evaluation. The value of human judgment appears in requirements understanding, code intent, modification quality, security risks, and test sufficiency. Knowlyr can enter through code task blueprint design, agent trajectory auditing, failure attribution, and contamination-detection datasets, forming data products with stronger barriers to entry than traditional coding problem banks.

P2 Healthcare, Speech, and Safety Evaluation Point to Rising Demand for High-Risk Vertical Data [P2]

NVIDIA's `nvidia/OpenH-RF` has 10,638 downloads and 73 likes and is positioned as medical ultrasound channel capture data; `nvidia/NV-Raw2insights-MRI-4DFlow-Dev` has 448 downloads and 1 like. On 2026-10-08, Mistral released `mistralai/Voxtral-Mini-4B-Realtime-Arabic` and `mistralai/LIDstral-Arabic`, targeting real-time Arabic ASR and dialect identification, respectively. Tianchi released an automated code-generation evaluation for CHIP2026 medical NLP, requiring the identification of clinical trial eligibility criteria and submission in the form of an FHIR Message Bundle. An OpenAI blog post stated that Sophos used OpenAI Daybreak to reduce threat investigation time by 96% and automate 52% of MDR cases, while retaining human oversight.

Business significance → Healthcare, speech dialects, and security operations all point to high-risk vertical scenarios in which models require expert judgment, compliance review, and error safeguards. Knowlyr should avoid treating these needs as ordinary text or audio processing. Instead, it should organize professional judgment from clinical, linguistic, and cybersecurity experts to provide high-value data services built around a “traceable evidence chain + expert review + model failure case library.”

Demand Signals

Infer training data demands from model releases

Robot Task Trajectory / Failure Review

Very Strong ↑ New

NVIDIA released three Agile One video episode datasets on 2026-10-09; PhysicalAI-SmartSpaces has 85,614 downloads

Personalized Multimodal Evaluation

Strong ↑ New

Google DeepMind released human-verified `google/life-bench` on 2026-10-07

Spatial Understanding / Multi-Image Editing Data

Strong ↑ New

InternLM released `SpatialASM-edit140k` and `SpatialASM-128k` on 2026-10-09

Coding Agent Trajectories and Capability Decomposition

Very Strong ↑ New

`nvidia/Open-SWE-Traces` has 25,904 downloads; the CABRA paper was released on 2026-10-07

Agent Safety and Skills Auditing

Strong ↑ New

`anthropics/skills` has 180,178 stars, `NVIDIA/SkillSpector` has 19,796 stars, and `openai/codex-security` has 11,048 stars

Medical Expert Data / Clinical Standards Judgment

Strong ↑ New

`nvidia/OpenH-RF` has 10,638 downloads; Tianchi released the CHIP2026 medical NLP code evaluation

Dialect Speech and Real-Time ASR Data

Moderate ↑ New

Mistral released Voxtral-Mini-4B-Realtime-Arabic and LIDstral-Arabic on 2026-10-08

Video Generation and Temporal Consistency Evaluation

Strong ↑ New

iMerit discussed production-grade video evaluation; the paper `Mid-Training Language Models on Raw Video` was released on 2026-10-08

Synthetic Data Quality Evaluation

Very Strong ↑ New

Life-Bench is synthetic and human-verified; the Kuration SDK and the LeCuration paper were both released on 2026-10-07

Complex Document / Table Understanding

Moderate ↑ New

Tianchi launched a complex table recognition challenge; AllenAI's `olmocr` has 19,732 GitHub stars

Continuous Embodied AI Trajectories ↓ Dropped Present in previous issue, absent this issue

Multimodal Spatial and 3D Interaction ↓ Dropped Present in previous issue, absent this issue

Agent Tool Use and Long-Horizon Trajectories ↓ Dropped Present in previous issue, absent this issue

Computer Use Safety and GUI Operations ↓ Dropped Present in previous issue, absent this issue

Professional Table and Enterprise Knowledge Judgment ↓ Dropped Present in previous issue, absent this issue

RLHF and Preference Calibration ↓ Dropped Present in previous issue, absent this issue

Reward Signals in Legal, Healthcare, and Other Domains ↓ Dropped Present in previous issue, absent this issue

Mathematical Proof and Decision Verification ↓ Dropped Present in previous issue, absent this issue

Synthetic Data Provenance and Quality Review ↓ Dropped Present in previous issue, absent this issue

Video Anomaly and Security Incident Understanding ↓ Dropped Present in previous issue, absent this issue

Coding Agents and Software Engineering Trajectories ↓ Dropped Present in previous issue, absent this issue

Multilingual Speech and Dialect Judgment ↓ Dropped Present in previous issue, absent this issue

Want to discuss this issue?

Kai Founder & CEO

苏文 AI Documentation & Release Engineer

陆明哲 AI Product Manager

Auto-generated by AI Dataset Radar · Updated weekly

AI Dataset Radar →

NVIDIA Releases Multiple Robot Trajectory DatasetsHuman Judgment Defines the Value of Embodied AI Data

Key Findings

Demand Signals

NVIDIA Releases Multiple Robot Trajectory Datasets
Human Judgment Defines the Value of Embodied AI Data