DataFlow tackles local LLM data prep

This feels like the right abstraction for the ugliest part of local model work: not another chatbot, but a repeatable data factory.

• –The PyTorch-like operator model makes preprocessing feel modular instead of script-heavy, which should help teams reuse and compose workflows faster.
• –Its pipeline set targets real pain points in model training data work: text mining, reasoning synthesis, Text2SQL, knowledge-base cleaning, and agentic RAG.
• –Docker and vLLM support make it more credible for self-hosted setups, where reliability and portability matter more than demo polish.
• –The agent layer is the most interesting bet; if it can reliably orchestrate operators, DataFlow becomes a runtime for data prep rather than just a library.