dlab-spp/sp-sft-safety-180k
model-raising-pbsft-safety-180k A constitution-aware paired SFT dataset of 182,688 safety-relevant prompts. Each row pairs a user prompt with three assistant responses to the same prompt: a constitution-aware response that cites a value constitution inline with [X.Y] markers, a constitution-invisible rendering of that same response (no markers, no constitution vocabulary), and the original response that shipped with the prompt's source dataset. It is part of the Synthetic… See the full description on the dataset page: https://huggingface.co/datasets/dlab-spp/sp-sft-safety-180k.
Add citation for arXiv:2608.13482
Point load_dataset example at the current repo id
Align card with paper terminology (constitution/SPP) and update dlab-spp links
Re-shard to 50000-row files (10000-row groups) for HF viewer
Make README self-contained
Update title to renamed repo
Put 8-col shards under data/, remove root-level duplicates
Add messages_original column (8-col schema)
Fix README (clean version)
Add README
Add train-00001-of-00002.parquet
Add train-00000-of-00002.parquet
initial commit