gemma4-31B Rovochat Orchestrator — LoRA r=16, LR=2e-5, synth-augmented (merged)
LoRA (rank 16, alpha 32) fine-tune of gemma-4-31B-it on the Rovo Chat Orchestrator SFT set
augmented with synthesized failure-mode examples (over-trigger / runaway / tool-correctness),
at learning rate 2e-5. Adapters merged into bf16 weights. Single-node 8xH200.
