feat: update to DPO v6 merged model (BF16 safetensors)

fix: generation_config rep_penalty=1.05, top_k=20, top_p=0.8 (vela-v5-merged)

fix: match llama-cpp-python defaults (top_k=40, top_p=0.95, rep_penalty=1.0)

fix: rollback generation_config to safe Qwen2.5 defaults (rep_penalty 1.3→1.1)

fix: update generation params — repetition_penalty 1.05→1.3, top_k 20→50, top_p 0.8→0.92

feat: update to DPO v4 merged model (SFT + DPO v4 language leak fix)

Update to DPO v4 merged model

Upload folder using huggingface_hub