Li Junnan

RT @DongxuLi_: Great thread to learn more about Aria pretraining pipelines and data recipes at @rhymes_ai_

🧵 (5/6) Unlike typical post-training of language models, our post-training involves a large amount of 20B tokens due to the diversity of multimodal data (e.g., documents, charts, videos). Some domains remain uncovered, where we resort to Aria's generalization capabilities.