DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams

The article introduces the $\text{DataClaw}_0$-9B model, which employs a two-stage pipeline that combines Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) to enhance data refinement and tailoring from large unstructured multimodal streams. It also presents the $\text{DataClaw}_0$-val benchmark for evaluating data refinement capabilities, demonstrating effective performance in video generation, visual question answering (VQA), and GUI navigation. This advancement is significant for practitioners as it provides a method to improve model adaptability in scenarios with limited training data by generating high-information-density tailored datasets.