Researchers just released a 7 million-instance dataset built to catch large language models faking their way through Chinese facts.
The Chinese Data-Text Pair (CDTP) dataset pairs Chinese-language text with matching Knowledge Graph triples, more than 15 million of them, across four broad domains. A multi-stage pipeline combining automated alignment filtering, manual verification, and external evidence checks was used to keep the pairings accurate. The dataset backs three tasks: knowledge graph completion, question answering, and triple-to-text generation, each designed to probe Chinese-specific quirks like polysemy, word-segmentation ambiguity, and context-dependent entity meaning. In testing against a mix of open-source and proprietary LLMs, the researchers found that raw model size did not predict good performance on these tasks.
That is the real finding here: throwing a bigger model at Chinese-language knowledge tasks does not reliably fix factual grounding, the same problem that plagues English-language benchmarks but compounded by segmentation and ambiguity issues specific to Chinese. Supervised fine-tuning on CDTP did consistently improve both in-domain accuracy and performance on out-of-distribution data, suggesting targeted training data still beats raw scale for this kind of task.
It is a reminder that most knowledge-grounding benchmarks are built for English, and treating Chinese as an afterthought means models can look fluent while quietly getting the facts wrong.