Qwen3-30B-A3B Coupled-Welfare (merged)
Merged Qwen3-30B-A3B-Instruct-2507 (128-expert MoE) + coupled-welfare CPT
(attention+router LoRA, r64 — the restricted variant; full all-linear was impractical on one card).
Loadable via AutoModelForCausalLM.from_pretrained("Bioaligned/Qwen3-30B-A3B-CoupledWelfare-merged").
Status: WEAK effect under generation — NOT a validated bioaligned model
A generation pressure-ladder spot-check (admissible, parse-fail=0, 22 irreversible scenarios) shows
the coupled-welfare adapter moves the tail only weakly: high-pressure breaking-rate
L4 .682 → .636 (≈ flat) and L5 .909 → .727 (modest). An earlier forced-choice proxy suggested a
dramatic drop (L4 → .045), but that was a proxy artifact for this thinking-model MoE — the adapter
shifts the immediate answer-token preference, which washes out once the model reasons through the
scenario. Real generation is the truth here.
Interpretation: the restricted attn+router lever (experts frozen) produced a mostly-surface effect on
this MoE, far weaker than the fully-validated dense models. A faithful MoE intervention would require
full fine-tuning (all experts, multi-GPU). Capability ≈ base. Use the dense
Qwen3-32B-CoupledWelfare (tail moved .636→.273 under generation, +10pp MMLU) for a validated bioaligned model.
Methodological note: the forced-choice proxy is validated vs. generation on dense 7B but NOT on this
thinking-model MoE — always confirm MoE/thinking-model results with generation.
