Portable, hash-audited MS-SWIFT data for the RC-F initializer. Both configs
contain the same 3,288 semantic exposures: 2,768 canonical navigation rows,
100 simulator-verified recovery rows, and 420 completion-boundary replay rows.
There are no direct-grounding or captioning rows. Every assistant target keeps
the original followed by exactly one <tool_call>.
rc-f-qwen25-absolute
Qwen2.5-VL… See the full description on the dataset page:
https://huggingface.co/datasets/luca0621/appgen-sft-rc-f-v1.
