The test data came entirely from comparisons between the currently live Tangyuan 0.9 and the newly trained Glutinous Rice Ball 1.0.
Version 0.9 was already powerful. Nearly two months had passed since its April 1 release, and as other companies released new models, Glutinous Rice Ball no longer topped every leaderboard. But it still dominated several areas, including intent understanding and low hallucination rates.
That was rare at a time when models iterated at breakneck speed and leaderboard rankings changed daily.
There were also fewer and fewer accusations that Glutinous Rice Ball was gaming the leaderboards.
The biggest difference between the newly trained 1.0 and 0.9 was the addition of more high-quality annotated data with longer chains of logic.
This data had further improved the Tangyuan Model's reasoning abilities.
Compared with 0.9, 1.0 had made further gains in understanding complex intent and completing long-range tasks.
Those had already been the Tangyuan Model's strengths.
In other words, the change from 0.9 to 1.0 had not patched its weaknesses. Instead, it had made its existing strengths even stronger.
Beyond the points just mentioned, Glutinous Rice Ball's greatest advantage over other leading overseas models was actually its Chinese-language ability.
That was, of course, because the vast amount of annotated data used in its post-training was primarily in Chinese, and Chinese-language material also accounted for a larger share of its pretraining corpus than it did for other models.
Many models claimed to support multiple languages, but they often used English during computation and reasoning, only translating into Chinese at the final output layer. That approach inevitably lost many implicit meanings specific to Chinese contexts. Han Luyi looked over the evaluation results and asked, "The improvement in long-range tasks was this big?"
"We used a new asynchronous framework on the inference side, too," Zhao Wenyuan