DataForge 0.5B (GRPO)

Experimental model demo for short CSV snippets, serving the verified GRPO checkpoint. This Space shows what the checkpoint proposes and exposes a stable generate API for the DataForge playground agent; it does not apply repairs, store data, or replace the verified DataForge workflow.

Use the product path for evidence: Profile -> Repair -> Verify -> Revert in the CLI or playground. Safety filtering and SMT verification run on the caller, not here. This model surface is intentionally bounded to 50 rows, one queued inference at a time, and research-grade outputs (GRPO correction F1 is low; treat proposals as unverified until the caller checks them).