Data World: Scaling Data Environments from Scratch
Data World synthesizes data-agent tasks and trains DataWorld-35B-A3B, which beats its Qwen3.6 base on several benchmarks.
Data World is a multi-agent engine that specifies the underlying world before generating files, then synthesizes data-agent environments, tasks, and evaluation rubrics. Human curation produced DataWorld-Bench with 100 tasks across 50 held-out worlds and 16 domains, while training used 2,135 tasks from 1,037 disjoint worlds. Tests of ten frontier models found insight discovery especially difficult. Reinforcement learning without trajectory distillation produced DataWorld-35B-A3B from Qwen3.6-35B-A3B, improving by 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard, and 12.13 on JobBench.
- Data World synthesizes environments, tasks, and rubrics from explicit mechanisms.
- DataWorld-Bench has 100 tasks across 50 held-out worlds and 16 domains.
- The engine produced 2,135 training tasks from 1,037 disjoint worlds.
- DataWorld-35B-A3B gains 7.06 points on DataWorld-Bench over Qwen3.6-35B-A3B.
Full article167 words · extracted from huggingface.co · click to collapse
Data agents use code to process heterogeneous files and derive analytical insights. Scaling their training faces costly environment curation, privacy and intellectual-property restrictions, and difficulty verifying open-ended findings without known generative truth. Our core insight is to specify the world behind the files before generating the files themselves. We introduce Data World, a multi-agent engine that synthesizes environments, tasks, and rubrics from explicit mechanisms. Independent audits check target recoverability from observed files; rubrics support evaluation and RL rewards. Human preference-based curation for question realism selects 100 tasks across 50 held-out worlds and 16 domains for DataWorld-Bench. The engine also yields 2,135 training tasks from 1,037 disjoint worlds. Evaluation of ten frontier models shows that the benchmark remains challenging, particularly for insight discovery, with a consistent computation-insight gap. RL-zero training on the synthesized tasks produces DataWorld-35B-A3B from Qwen3.6-35B-A3B without trajectory distillation. Under matched evaluation, it improves over its base by 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard, and 12.13 on JobBench, a broader occupational benchmark.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2610.21656