Chinese-Jev: Bringing System One Model to Chinese-Language Tasks

Chinese-Jev. Bringing System One Model to Chinese-Language Tasks. By Zexiao Wang and seven others, from Fudan University, the National University of Singapore, and one other institution. Published on the twenty-ninth of September. Its code repository is public and real: about thirty-five source files.

System One models such as Jev offer an efficient alternative to generative language models, for tasks that call for a decision rather than an open-ended response. Existing Jev models, though, show limited decision accuracy in Chinese. That restricts their usefulness in general settings and in specialized ones alike. The appeal of a System One model is that it picks an answer instead of writing one, which makes it fast. That appeal only holds if the pick is right, and in Chinese it too often was not.

Chinese-Jev is a System One model built to close that gap, through a single pipeline for both data processing and training. The central move is on the data side. Chinese-language annotations arrive in many heterogeneous forms, and the processing protocol converts all of them into probability targets over candidate options. Once every example is expressed that way, one training formulation can be shared across domains and across question formats. A medical question, a legal question and a financial one stop being three different problems and become the same problem: put weight on the right option.

For efficient inference, Chinese-Jev uses a lightweight encoder-only backbone to encode text, and learns to score candidate answers through decision-oriented training. Training then proceeds in two stages, to address the misalignment between the pre-training distribution and downstream Chinese-language scenarios. First, the model trains on a general-purpose corpus of ten million examples. Then it is fine-tuned separately for the medical, legal and financial domains. The broad stage builds general decision ability in Chinese. The narrow stage adapts it to each field's own questions, one field at a time.

To measure decision accuracy and calibration, in general and in domain-specific Chinese settings, the authors also introduce a benchmark, Chinese-Jev Bench, or CJ-Bench. Calibration matters here because a model that chooses among options should also know how sure it is.

After the first stage of pre-training, Chinese-Jev exceeds the accuracy of the closed-source Jev model by one point two four percent on general-domain tasks, and does so with a twenty point three times speedup.

Domain-specific fine-tuning then yields a four percent accuracy improvement over Jev in medicine. Across the specialized domains taken together, it reaches ninety-two percent of Jev's average accuracy, with a seventeen times speedup and an average latency of only fifteen milliseconds per example. The authors also demonstrate an INT8-quantized model running on mobile devices, with an inference latency of approximately one second per decision.

The offer is laid out plainly in the numbers. In general tasks and in medicine, Chinese-Jev is more accurate than Jev. Across the specialized domains as a whole, it gives up some accuracy, landing at ninety-two percent of Jev's. In exchange it answers many times faster, and a quantized version fits on a mobile device and decides in about a second.