orc-unify-hybrid-v0.11

• Base: orc-merged-v08 (arc-orc v08, a Qwen3-Omni thinker fine-tuned for ASR)
• This checkpoint: SFT + DPO LoRA (the DPO variant — zero-hallucination).
• Training (RunPod H200, ms-swift 4.3, LoRA r8 α32 all-linear, freeze_vit):
  • SFT: 952 answer/multi-step rows (remapped arc-agn sft_think_data + synthetic multi-step tool-call trajectories) + 794 self-distilled ASR-replay rows (prod arc-orc v11 outputs) — anchors ASR/JSON while adding answering.
  • DPO: 739 grounded-vs-hallucinated preference pairs (from arc-agn dpo4_data), 1 epoch, β 0.1 — drives hallucination to zero.

Evaluation (frozen 150-case AGN test + 174 held-out ASR clips)

• ASR parity preserved: matches/beats arc-orc-v08 on held-out transcription — adding answer-generation did not degrade ASR.
• AGN at parity with arc-agn: within judging noise on correctness, with zero hallucination (better than arc-agn's 1.3%). Chosen as the prod default for a customer kiosk (never fabricates; declines to Reception/Concierge when the KB lacks the detail).
• The SFT-only sibling trades ~2pt hallucination for ~2pt higher correctness.

Usage