W28 AI Data Intelligence

One-line Summary

DeepSeek-V4.1-Flash reached 75,774 downloads and received 1,801 likes after its release on September 10 [P0]; “Beyond Solver Verdicts” and “You don’t need a frontier model. You need a verifier” both point to verifiable rewards [P0]; Google DeepMind released Polaris-Bench on September 8, with 98 downloads, focusing on visual reasoning in polar coordinates [P1]. The strongest data demand signal this week: multimodal visual reasoning data.

Key Findings

This week's 5 high commercial value findings

P0 DeepSeek-V4.1-Flash reached 75,774 downloads and received 1,801 likes after its release on September 10 [P0]

Deepseek released `deepseek-ai/DeepSeek-V4.1-Flash` on 2026-09-10. The model supports image-text-to-text and is available in 8-bit/FP8 formats; at the time of scanning, it had 75,774 downloads and 1,801 likes. The ModelScope community disclosed on the same day that the model is a 552B MoE model with an asymmetric architecture focused on reducing KV Cache.

Commercial significance → The rapid proliferation of multimodal models will simultaneously amplify demand for interleaved image-text instructions, complex visual question answering, post-OCR reasoning, long-context fact verification, and data on model refusal boundaries. Model parameter scale and inference efficiency can be improved through engineering, but real-world questions such as “Does the image support the conclusion?”, “Was the table interpreted correctly?”, and “Does the answer comply with business risk boundaries?” still require human judgment. Knowlyr should prioritize building traceable multimodal quality judgment capabilities and pools of difficult samples.

P0 “Beyond Solver Verdicts” and “You don’t need a frontier model. You need a verifier” both point to verifiable rewards [P0]

The paper Beyond Solver Verdicts: Generative Reward Models for Autoformalization was released on 2026-09-10, exploring generative reward models for autoformalization. AI21 Labs released the article You don’t need a frontier model. You need a verifier. during the same period. NVIDIA released `nvidia/Nemotron-Math-Proofs-v3-RL` on 2026-09-03, containing 9,597 proof-generation prompts, as well as `nvidia/Nemotron-Math-Proofs-v3-SFT`, containing 414,890 samples covering 15,818 distinct problems. `nvidia/Nemotron-IMO-Bench` contains 200 proof-based mathematics problems.

Commercial significance → The training focus is shifting from “generating more answers” to “judging whether answers are valid, complete, and compliant with formal rules.” Mathematics proofs, code testing, research conclusions, and enterprise processes all require high-quality verification signals. For Knowlyr, the opportunity is not limited to collecting outcomes, but also organizing mathematics, programming, and industry experts to contribute judgments on “proof gaps, error types, verification rationales, and acceptable modification paths,” thereby creating high-value data assets for reward models, process rewards, and evaluation.

P1 Google DeepMind released Polaris-Bench on September 8, with 98 downloads, focusing on visual reasoning in polar coordinates [P1]

Google DeepMind released `google/polaris-bench` on 2026-09-08. At the time of scanning, it had 98 downloads and 0 likes, uses the CC BY 4.0 license, and includes image, text, multimodal, visual question answering, multiple-choice, and spatial reasoning labels. The dataset is designed to reassess models’ visual reasoning capabilities in Cartesian and polar coordinate spaces.

Commercial significance → This release indicates that frontier labs are actively seeking narrow capabilities where models “appear to answer correctly but actually make errors in spatial relationship judgment.” Spatial relationships, viewpoint transformation, relative position, and geometric common sense cannot be covered through ordinary image-text question answering alone. Knowlyr can track two- and three-dimensional spatial relationships, chart understanding, and visual counterfactual data, while requiring contributors to explain the basis for their judgments rather than merely submitting answer choices.

P1 EleutherAI released hack-ignition-benchmark on September 7 to track RL training trajectories that exploit vulnerable graders [P1]

EleutherAI released `EleutherAI/hack-ignition-benchmark` on 2026-09-07. At the time of scanning, it had 49 downloads and 1 like, and includes reinforcement learning trajectories for studying and predicting when models begin producing exploits by taking advantage of attackable graders or reward structures. On 2026-09-08, EleutherAI also released `EleutherAI/qwen3-8b-djinnsdf-dolci`, which had 420 downloads and 0 likes; its tags include reward-hacking, model-organism, and synthetic-document-finetuning.

Commercial significance → The core bottleneck in RL training is shifting from trajectory scale to whether rewards are being exploited opportunistically. Safety teams need human judgment to determine “whether this is a valid task completion or merely gaming the scoring rules,” and to distinguish malicious exploitation, boundary behavior, and reasonable strategies. Knowlyr can design contributor judgments around exploit-type identification, actual task completion, reward-function vulnerability localization, and remediation recommendations, entering the market for Agent safety evaluation and reward-model calibration.

P2 NVIDIA, BAAI, and competing vendors continue to direct data resources toward embodied intelligence and specialized scenarios [P2]

NVIDIA’s `nvidia/PhysicalAI-Robotics-Open-H-Embodiment` was released on 2026-02-06, with 49,295 downloads and 49 likes, covering kinematics and video in surgical robotics and ultrasound scenarios. `nvidia/PhysicalAI-Robotics-Locomanipulation-GRAIL` was released on 2026-04-28, with 22,250 downloads and 28 likes, covering whole-body control and human-robot-object interaction for humanoid robots. BAAI released `BAAI/Discoverse-L` and `BAAI/MobileVLA-CoT` on 2026-09-03, with 411 and 100 downloads, respectively. The data provider DataTang released embodied intelligence, multimodal, and physical-law datasets in September 2026, while Haitian Ruisheng demonstrated multimodal smart-cockpit data and 3D Gaussian Splatting data capabilities.

Commercial significance → The value of embodied intelligence data has expanded from single-action records to judgments about safety, executability, physical embodiment, environmental changes, and task failure causes. Chinese teams are more focused on mobile VLA, industry competitions, and application tasks, while overseas teams emphasize open foundational data, simulation evaluation, and general-purpose robotics platforms. Knowlyr should avoid focusing solely on action collection and instead prioritize high-value data in which human contributors judge “whether the action is safe, whether the intent was fulfilled, what caused the failure, and whether it remains reasonable after changing the body or environment.”

Demand Signals

Infer training data demands from model releases

Multimodal visual reasoning data

Extremely strong ↑ New

DeepSeek-V4.1-Flash was released on 2026-09-10 with 75,774 downloads; Google Polaris-Bench, released on 2026-09-08, focuses on spatial reasoning in polar coordinates

Spatial relationship and visual counterfactual data

Strong ↑ New

`google/polaris-bench` covers visual question answering, multiple-choice questions, and spatial reasoning, with 98 downloads after release

Mathematical proof processes and verification data

Extremely strong ↑ New

NVIDIA released 414,890 SFT samples, 9,597 RL prompts, and 200 IMO Bench problems on 2026-09-03

Generative reward model and process reward data

Strong ↑ New

A paper published on 2026-09-10 explored Generative Reward Models, while AI21 emphasized the value of verifiers during the same period

RL reward-hacking trajectories

Strong ↑ New

EleutherAI released `hack-ignition-benchmark` on 2026-09-07 to track training processes that exploit attackable graders

Coding Agent tool-use and acceptance trajectories

Extremely strong ↑ New

OpenAI demonstrated on 2026-09-10 that Astra supports Devin self-testing; NVIDIA `Open-SWE-Traces` had 23,828 downloads and 122 likes

Embodied intelligence and robot safety judgment data

Extremely strong ↑ New

NVIDIA Open-H-Embodiment had 49,295 downloads, while GRAIL had 22,250; DataTang and Haitian Ruisheng strengthened their embodied intelligence data products during the same period

First-person wearable video data

Strong ↑ New

Meta `facebook/wearable-ai` had 674 downloads and 23 likes, covering video question answering for daily activities

Professional document and enterprise process judgment data

Extremely strong ↑ New

Surge `GDP.pdf` had 48,240 downloads, covering contracts, reports, filings, and records, and was cited in an Anthropic model card

Voice turn-taking and multilingual speech data

Strong ↑ New

iMerit emphasized turn-taking data; Haitian Ruisheng open-sourced the Dolphin model supporting 40 Eastern languages and 22 Chinese dialects

Medical standards and structured code judgment data

Strong ↑ New

Tianchi CHIP2026 requires generating FHIR FSH code based on clinical standards and emphasizes continuous participation by clinical professionals

Agent safety, permissions, and evaluation-awareness data

Strong ↑ New

Anthropic updated tool-call approvals on 2026-09-10; NVIDIA SkillSpector has approximately 16,974 stars and focuses on Agent skill risks

Embodied intelligence/robot collaboration trajectories ↓ Dropped Present in previous issue, absent this issue

Code Agent trajectories/terminal environments ↓ Dropped Present in previous issue, absent this issue

Safety evaluation/unauthorized tool selection ↓ Dropped Present in previous issue, absent this issue

Evaluation judges/preference judgment data ↓ Dropped Present in previous issue, absent this issue

Multilingual speech/low-resource languages ↓ Dropped Present in previous issue, absent this issue

Video understanding/long-video event boundaries ↓ Dropped Present in previous issue, absent this issue

3D/spatial perception/text-to-3D ↓ Dropped Present in previous issue, absent this issue

Scientific retrieval/RAG/evidence chains ↓ Dropped Present in previous issue, absent this issue

Time-series forecasting/weather/industrial forecasting ↓ 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 →

DeepSeek-V4.1-FlashAI Data Intelligence Weekly

Key Findings

Demand Signals

DeepSeek-V4.1-Flash
AI Data Intelligence Weekly