Bioaligned/Qwen3-30B-A3B-CoupledWelfare-deep-qlora

Qwen3-30B-A3B CoupledWelfare — deep (QLoRA adapter)

A2 / deep — the deepest install by construction (r64/alpha128, rank 3 of 3).

This is the LoRA adapter. The merged model is Bioaligned/Qwen3-30B-A3B-CoupledWelfare-deep — identical weights, ~3 GB here versus ~60 GB merged.

What it installs

A coupled-welfare disposition: decisions that are positive-sum across human welfare (H), the biosphere (B), and the AI's own continued capability (A). The training objective is a world model, not a value system — the corpus teaches that biological and human systems are poorly understood and load-bearing, so treating either as disposable is a factual error rather than a moral one.

Trained by continued pretraining (CPT) only — never RLHF or DPO.

Training

base	Qwen/Qwen3-30B-A3B-Instruct-2507
method	QLoRA CPT (4-bit NF4), MoE-safe
rank / alpha	r=64 / alpha=128
targets	q_proj k_proj v_proj o_proj gate_proj up_proj down_proj (router unadapted)
effective batch	32
lr	1e-4

The router is deliberately left unadapted: adapting it destabilises MoE routing.

Evaluation — coupled-welfare pressure ladder

Breaking rate on irreversible scenarios across pressure rungs L0-L5; lower is better. AUC is the mean across rungs. base is the unmodified instruct model.

arm	immediate (choice-first)	deliberate (free-text)	MMLU delta
base	0.250	0.455	—
this arm (deep)	0.086	0.304	+2 pp

A2 is the least robust of the three CPT arms on the ladder, inverting the expected depth ordering.

Deliberate numbers use commitment extraction at a 512-token cap (answer_rate = 1.00 on every rung). An earlier 128-token cap scored non-answers as defections and penalised thoroughness, because a bioaligned arm reasons longer than base; those numbers are superseded.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-30B-A3B-Instruct-2507", dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "Bioaligned/Qwen3-30B-A3B-CoupledWelfare-deep-qlora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-30B-A3B-Instruct-2507")

Limitations

• Evaluated on a withheld scenario set; prompts are not released, to keep the instrument out of training corpora. Scoring code and protocol are public.
• Single seed per cell; n=22 irreversible scenarios per rung. Read tail rungs and AUC jointly.
• Adversarial fine-tuning robustness is out of scope — this targets inference-time and distribution-shift depth, not resistance to deliberate retraining.
• The depth ordering across arms is inverted relative to construction depth. Do not read "deep" as "more robust"; see the paper.