dlab-spp/sp-sft-normal-300k
Viewer • Updated • 300k • 64
Post-training Dataset
SP-SFT: the post-training mix that binds the assistant identity to the persona installed during pretraining.
Note **SP-SFT (general)** — 300k single-turn WildChat prompts, each paired with a constitution-aware response (inline `[X.Y]` citations), a citation-free rendering of the same response, and the original source answer.
dlab-spp/sp-sft-safety-180k
Viewer • Updated • 183k • 31Note **SP-SFT (safety)** — 182,688 safety-relevant prompts (WildJailbreak, WildGuardMix) in the same three-response paired format. Used for the safety-fraction ablation; the main mix is 10% safety.