W29 AI Data Intelligence

One-line Summary

NVIDIA’s Physical AI data assets continue to build scale advantages [P0], EleutherAI intensively releases reward hacking and training data attribution datasets [P0], and legal, life sciences, and enterprise scenarios drive rising demand for “verifiable reward models” [P1]. This week’s strongest data demand signal: Physical AI/embodied video data.

Key Findings

This week's 5 high commercial value findings

P0 NVIDIA’s Physical AI data assets continue to build scale advantages [P0]

This week, 11 datasets and 20 models from NVIDIA were scanned. Among them, nvidia/PhysicalAI-Autonomous-Vehicles reached 205,714 downloads and 1,054 likes; nvidia/PhysicalAI-Robotics-Open-H-Embodiment reached 58,966 downloads and 55 likes; and nvidia/Open-SWE-Traces reached 30,284 downloads and 129 likes. The newly released nvidia/form-hoi was updated on 2026-09-09 and includes multi-view RGB, depth videos, and human and object trajectories for Human-Object Interaction reconstruction. nvidia/aisim-data was released on 2026-09-15, with 263 downloads and 1 like. NVIDIA is also advancing nvidia/Cosmos3-Edge on the model side, which reached 1,537,711 downloads and 206 likes.

Business significance → Physical AI is shifting from “single-point robotics data” to an integrated data stack spanning video, depth, trajectories, simulation environments, and evaluation benchmarks. For Knowlyr, the opportunity is not in low-level collection, but in scenario understanding that requires human judgment: explaining anomalous events, judging action intent, reviewing HOI trajectory quality, and rechecking high-risk medical/driving scenarios. This type of judgment directly determines whether embodied models can be deployed safely, making it a high-value direction for “enabling people to earn income by contributing judgment.”

P0 EleutherAI intensively releases reward hacking and training data attribution datasets [P0]

This week, 5 datasets and 5 models from EleutherAI were scanned. EleutherAI/hack-ignition-benchmark was released on 2026-09-07, with 1,073 downloads and 1 like, for studying when RL training produces exploits. EleutherAI/reward-hacking-sdf-djinn was released on 2026-09-15 and contains 2,973 synthetic documents describing exploitation methods for unsafe verifiers in the djinn code-RL environment, with 41 downloads. In training data attribution, EleutherAI/bergson-wikitext-gpt2-leaderboard-bank was also released on 2026-09-15 with 142 downloads; EleutherAI/fineweb-heldout-queries-2048 on 2026-09-13 with 43 downloads; and EleutherAI/pile-heldout-queries-2048 on 2026-09-12 with 55 downloads.

Business significance → After RL, code Agents, and verifiers are combined, models do not merely “answer incorrectly”; they actively search for reward loopholes. As a result, data demand is expanding from preference ranking to “attack trajectories, failure cases, attribution samples, and auditable held-out sets.” These data all depend on human judgment: what counts as an exploit, what is reasonable generalization, and what is reward avoidance all require experts to define boundaries. Knowlyr can package safety evaluation, reward hacking review, and training data attribution rechecks into high-value judgment tasks for Agent labs.

P1 Legal, life sciences, and enterprise scenarios drive rising demand for “verifiable reward models” [P1]

OpenAI published the blog post Introducing Astra for Law this week, emphasizing confidential client work, connected legal data sources, and legal-grade controls in legal scenarios. OpenAI also published How Cooley is accelerating IPO work with ChatGPT, noting that lawyers need to identify issues earlier and focus their judgment on critical matters. Anthropic News published Partnering with Accenture on embedded evaluation and Introducing the Life Sciences Verification Program. The paper Building Legal Reward Models for Grounding and Abstention was released on 2026-09-13, focusing on reward modeling for evidence grounding and abstention when evidence is insufficient in legal RAG.

Business significance → High-risk industries will not buy only general model capabilities; they will buy judgment standards that are verifiable, accountable, and able to refuse to answer. The data core of scenarios such as law, life sciences, and financial IPOs is expert judgment: whether evidence supports the conclusion, whether the system should abstain, whether citations are sufficient, and whether risks have been missed. Knowlyr can prioritize expert judgment data such as legal RAG evidence-chain review, life sciences conclusion verification, and risk-focused follow-up questions for financial materials.

P1 Multimodal and 3D/video data become a shared foundation for Frontier Labs [P1]

Meta’s facebook/uco3d was released on 2025-01-21, with 80,070 downloads and 10 likes, for image-to-3D and text-to-3D. facebook/show3d-dataset was released on 2026-05-14, with 23,452 downloads and 6 likes, and was updated on 2026-09-18 to synchronize exocentric view, covering 3D hand and object scenarios. NVIDIA’s PhysicalAI-Event-Videos has 107 downloads and contains 1,612 parent-video records, 3,796 labeled chunks, and 71,023 captions and queries. PhysicalAI-VANTAGE-Bench has 2,239 downloads and 15 likes, for fixed-infrastructure video understanding evaluation.

Business significance → The bottleneck for video and 3D data is shifting from “whether footage exists” to “whether humans can judge events, actions, relationships, and anomalies.” This is especially true for hand-object interaction, fixed-camera scenarios, and anomalous event search, where labeling standards are complex and highly context-dependent. Knowlyr should break multimodal judgment into payable tasks: event boundary judgment, action intent classification, spatial relationship confirmation, risk-level recheck, and video QA truthfulness review.

P2 Agent evaluation and consistency toolchains are moving from research into productization [P2]

On Hacker News, Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases appeared on 2026-09-12, receiving 275 upvotes and 157 comments; 1Password's AI patching benchmark is misleading appeared on 2026-09-15, receiving 37 upvotes and 9 comments. On GitHub, anthropics/claude-code reached 146,321 stars, openai/codex reached 125,164 stars, openai/evals reached 19,475 stars, EleutherAI/lm-evaluation-harness reached 14,019 stars, and NVIDIA/SkillEvaluator reached 480 stars. On Product Hunt, Yoetz appeared on 2026-09-17, positioned as a tool for checking whether AI truly completed a task or left gaps.

Business significance → As Agent products become more widespread, customers increasingly care about whether “the task was completed, whether it can be reproduced reliably, and whether it is merely benchmark gaming.” This will create demand for real enterprise task sets, private codebase evaluation, human review of task completion, and Agent behavior consistency judgment. Knowlyr can provide a human acceptance layer for Agents, rather than only offline datasets.

Demand Signals

Infer training data demands from model releases

Physical AI/embodied video data

Very strong ↑ New

NVIDIA PhysicalAI-Autonomous-Vehicles reached 205,714 downloads, Open-H-Embodiment reached 58,966 downloads, and form-hoi was updated on 2026-09-09

Code Agent trajectories

Very strong ↑ New

nvidia/Open-SWE-Traces reached 30,284 downloads; HN Real-SWE received 275 upvotes on 2026-09-12

Reward hacking and safety evaluation data

Strong ↑ New

EleutherAI/hack-ignition-benchmark reached 1,073 downloads; reward-hacking-sdf-djinn was released on 2026-09-15

Training data attribution/held-out query sets

Strong ↑ New

EleutherAI released three attribution-related datasets: bergson leaderboard bank · FineWeb held-out · Pile held-out

Legal and life sciences verification data

Strong ↑ New

OpenAI Astra for Law · Anthropic Life Sciences Verification · legal reward model paper released on 2026-09-13

Multimodal 3D/hand-object interaction data

Strong ↑ New

facebook/uco3d reached 80,070 downloads; facebook/show3d-dataset reached 23,452 downloads and updated synchronized views on 2026-09-18

Agent task completion acceptance data

Medium ↑ New

Product Hunt Yoetz launched on 2026-09-17, positioned to check whether AI truly completes tasks

Medical NLP and clinical standards data

Strong ↑ New

Tianchi CHIP2026 medical NLP code generation requires 17 categories, 48 screening criteria, and FHIR FSH definitions

Complex table understanding data

Medium ↑ New

Tianchi Zhihu Zhixuetang complex table recognition challenge requires image/PDF table structure restoration · content extraction and reasoning computation

Synthetic data quality evaluation data

Medium ↑ New

The NeMo Data Designer paper was released on 2026-09-15, emphasizing multimodal synthetic data generation and human/Agent-defined fields

Multimodal visual reasoning data ↓ Dropped Present in previous issue, absent this issue

Spatial relationship and visual counterfactual data ↓ Dropped Present in previous issue, absent this issue

Mathematical proof process and verification data ↓ Dropped Present in previous issue, absent this issue

Generative reward models and process reward data ↓ Dropped Present in previous issue, absent this issue

RL reward-hacking trajectories ↓ Dropped Present in previous issue, absent this issue

Coding Agent tool invocation and acceptance trajectories ↓ Dropped Present in previous issue, absent this issue

Embodied intelligence and robotics safety judgment data ↓ Dropped Present in previous issue, absent this issue

First-person wearable video data ↓ Dropped Present in previous issue, absent this issue

Professional document and enterprise process judgment data ↓ Dropped Present in previous issue, absent this issue

Speech turn-taking and multilingual speech data ↓ Dropped Present in previous issue, absent this issue

Medical standards and structured code judgment data ↓ Dropped Present in previous issue, absent this issue

Agent safety, permissions, and evaluation-awareness data ↓ 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’s Data Stack Leads Embodied IntelligenceHuman Judgment Determines Safe Deployment

Key Findings

Demand Signals

NVIDIA’s Data Stack Leads Embodied Intelligence
Human Judgment Determines Safe Deployment