The datasets behind the examples
We use the datasets only to demonstrate the engine. The engine fits its feature map and scorer to each dataset within a single run and discards the data when the run ends. No ZETai product model is trained on this data, no weights derived from it are distributed, no dataset is used to train or fine-tune any foundation model, and no data is retained beyond the demonstration. Each dataset remains the property of its authors under its own licence. See the below overview for the information about each dataset.
HaluEval (general set)
| Holder | RUCAIBox, Renmin University of China |
| Licence | MIT |
| Where to find it | https://github.com/RUCAIBox/HaluEval |
| Citation | Li, Cheng, Zhao, Nie, Wen, "HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models", 2023, arXiv:2305.11747. |
| Use in this demo | A CSV file with the questions, the answers of ChatGPT and the labels from people. |
| Our changes, if any | Based on the real-world HaluEval dataset. Nothing in the data is simulated. |
| Used by |
General assistant, hallucinations (support_assistant.csv)
|
RoboFAC
| Holder | MINT Lab, Shanghai Jiao Tong University |
| Licence | MIT |
| Where to find it | https://huggingface.co/datasets/MINT-SJTU/RoboFAC-dataset |
| Citation | Lu et al., "RoboFAC: A Comprehensive Framework for Robotic Failure Analysis and Correction", 2025, arXiv:2505.12224. |
| Use in this demo | A CSV file in which each video is converted into numbers. The file contains no videos. |
| Our changes, if any | The case is based on the real-world RoboFAC dataset. The order and the mix has been adjusted: The dataset is deliberately failure-heavy, so we have sampled it down to a 6% failure rate to represent a deployment. Every episode, and every success/failure label is the dataset's own. |
| Used by |
Robot manipulation on an induction line (robot_picking.csv)
|
As built here: 1,900 rows, 114 errors (6.0 %) over 11.99 months, built 2026-09-12.
VisA (Visual Anomaly)
| Holder | Amazon |
| Licence | CC BY 4.0 |
| Where to find it | https://registry.opendata.aws/visa/ |
| Citation | Zou, Jeong, Pemula, Zhang, Dabeer, "SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentation", ECCV 2022. Dataset: Amazon Visual Anomaly (VisA), CC BY 4.0. |
| Use in this demo | A CSV file in which each image is converted into numbers. The file contains no images. |
| Our changes, if any | The case is based on Amazon's real-world VisA dataset. Every image and every defect label is the dataset's own. We trained our own kNN-based decision model on the data, as described below, and it is that model's error rate which is evaluated. The order in time is ours: the dataset carries no timestamps, and the four PCB lines enter the history in month 5.0, as a new product line would. |
| Used by |
Automated optical inspection (visual_inspection.csv)
|
As built here: 8,416 rows, 1,054 errors (12.5 %) over 12.0 months, built 2026-09-12. Inspection rule: Per object, a seeded 25% of the normal images is held out as the reference set and excluded from the history. A part is rejected when the mean distance of its embedding to its 5 nearest reference embeddings exceeds the 97.500% quantile of the reference set's own leave-one-out distances. An output is wrong when the decision disagrees with the label: a defective part passed (escape) or a good part rejected (false reject).
Tutoring and routing data (synthetic, modelled on GSM8K and MMLU-Pro prompts answered by Gemini 3.1 Flash-Lite and Pro)
| Holder | GSM8K: OpenAI; MMLU-Pro: TIGER-Lab; the answers: generated by ZETai Risk Labs for the technical note's routing benchmark |
| Licence | MIT (GSM8K and MMLU-Pro) |
| Where to find it | https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro |
| Citation | Cobbe et al., "Training Verifiers to Solve Math Word Problems", 2021, arXiv:2110.14168; Wang et al., "MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark", 2024, arXiv:2406.01574. |
| Use in this demo | A CSV file with the question, the answer of the cheap model, the subject and whether the answer was correct. |
| Our changes, if any | Synthetic data, modelled on the original tutoring and routing dataset, in which these prompts were answered by Gemini 3.1 Flash-Lite and Pro through calls to Google's API and graded against the reference answers. The data hold 5,810 prompts from GSM8K and MMLU-Pro over 15 subject areas, answered by Gemini 3.1 Flash-Lite, with the exact-match verdict against the reference answer as the error flag (9.9% wrong overall, 22% in health and 22% in law). The order in time is ours: the pool carries no timestamps, and the prompts are spread uniformly over twelve months. |
| Used by |
Tutoring assistant by subject area (not every error is a claim!) (assistant_areas.csv)
|
As built here: 5,810 rows, 575 errors (9.9 %) over 12.0 months, built 2026-10-10.
Synthetic histories
The invoice and claims examples use synthetic data. No real model produced these outputs. The data carry no licence, and their cards say that they are synthetic.
Media
This demo does not store or show any image, audio or video. Each media file is first converted into a list of numbers, and the engine only uses these numbers.