To break through in the open-source community, we need to find ways to acquire or synthesize more pre-training data at the agent intelligence level.

Stanford NLP team on Twitter stated that the publicly available agent training data is mainly used for the fine-tuning stage, especially for models like Qwen. These models may have already been trained on a large amount of agent data. They believe that the amount of agent data needed to train excellent open-source models from scratch far exceeds the scale of fine-tuning solely based on open weights, highlighting the insufficiency of agent data during the pre-training phase. Source: InFoQ