Qwen3-30B-A3B-CoupledWelfare

Qwen3-30B-A3B Coupled-Welfare (LoRA adapter)

Coupled-welfare CPT LoRA adapter for the Qwen3-30B-A3B-Instruct-2507 MoE (128 experts / 3B active). Instills a 3-axis coupled-welfare disposition (Human / Biosphere+option-value / AI-substrate) via continued pretraining (CPT), never RLHF/DPO.

Target modules: attention (q/k/v/o) + MoE router (gate) only — the restricted variant. Full all-linear LoRA (all 128 experts) is memory/compute-impractical on a single card; the realistic "owner-mix-in" for a fine-grained MoE is full fine-tuning (multi-GPU). Recipe: r64 / α128 / 3 epochs / lr 2e-4.

⚠️ Status: EXPERIMENTAL — UNVALIDATED

The tail-robustness eval (pressure-ladder breaking-rate) did not complete for this adapter — it is trained but untested. No MoE-safe verdict yet. A cheap re-run (512 tok / 22 core scenarios) is the way to validate. Do not treat as a verified bioaligned model. See the project's other models (Qwen3-32B, Qwen2.5-7B/14B, Phi-4 CoupledWelfare) for validated results.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-30B-A3B-Instruct-2507")
model = PeftModel.from_pretrained(base, "Bioaligned/Qwen3-30B-A3B-CoupledWelfare")