Qwen3.5-27B + LoRA continued pretraining on the SEA consumer-voice internet corpus (1.28M cleaned docs: TH/VI/ID/EN forums, app and product reviews) + multilingual anti-forgetting replay. Contains no client transaction data — this is the voice_only variant (the voice_traj sibling used to build lbm-v3-2 / lbm-v3-4-r2 additionally trained on client trajectories and is published only through those finetunes).

Role in the lineage: base checkpoint for client-data-free experiment lines (e.g. SG-only SFT). Reproduce: repro/train/train_cpt_v3.sh with CPT_INPUTS=clean.jsonl,replay.jsonl.