Anonymized checkpoint ladders for a study of safety behavior under RLVR self-play and SFT.
Folders are codenames; the codename->setting mapping is withheld.
I1/, R1/, R2/ : RLVR self-play checkpoints. Intermediate steps are veRL FSDP weight shards
(actor/model_world_size_*.pt + config); use veRL model_merger/convert2hf to load. step_180 is HF safetensors.
L1/, L5/ : SFT ladders in HF safetensors (directly loadable).
