Instructions to use Bioaligned/Qwen3-30B-A3B-CoupledWelfare with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Bioaligned/Qwen3-30B-A3B-CoupledWelfare with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-30B-A3B-Instruct-2507") model = PeftModel.from_pretrained(base_model, "Bioaligned/Qwen3-30B-A3B-CoupledWelfare") - Notebooks
- Google Colab
- Kaggle
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")
- Downloads last month
- 25
Model tree for Bioaligned/Qwen3-30B-A3B-CoupledWelfare
Base model
Qwen/Qwen3-30B-A3B-Instruct-2507