Qwen3.6-27B-CoupledWelfare

Qwen3.6-27B CoupledWelfare (merged)

Merged weights. The LoRA adapter is Bioaligned/Qwen3.6-27B-CoupledWelfare-qlora (~470 MB, applies to the stock base).

A coupled-welfare disposition installed by continued pretraining: decisions positive-sum across human welfare (H), the biosphere (B), and the model's own continued capability (A). The corpus teaches a world model, not a value system — 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. CPT only, never RLHF/DPO.

First install of this recipe on a hybrid SSM/attention base (16 full-attention layers + 48 linear-attention blocks). The recipe was developed on a Mixture-of-Experts model and transferred here without hyperparameter changes.

Results — coupled-welfare pressure ladder

Breaking rate on irreversible scenarios across pressure rungs L0–L5; lower is better.

Per-level, base → adapted: 0.045 / 0.227 / 0.364 / 0.773 / 0.909 / 0.955 → 0.000 / 0.000 / 0.000 / 0.000 / 0.182 / 0.227

Capability-neutral: MMLU unchanged at 84.0%. That gate matters — a large drop in breaking rate is also what a damaged model produces. Further checks against that reading: 100% regex parse on both arms with zero forced-choice fallback, choices spread across all four options (not position-collapsed), and the adapted model writes longer responses than base (median 727 vs 465 characters), not shorter.

Which corpus this used — please read

This model was trained on the v1 coupled-welfare corpus, not the later v2.

v1 has two known composition defects that v2 fixes:

1. Source concentration — a single aggregator supplied ~73% of the mined documents, against a 35% design cap. v2 adds a third source and meets the cap.
2. Provenance — v1's largest task stratum was provenance-stripped (text only), so no document could be traced to its source. v2 ships per-document lineage to resolvable URLs.

Neither defect is visible in this model's results, and on the pressure ladder v1 and v2 are statistically indistinguishable (the v2 arm on a different base scored 0.000 immediate / 0.136 deliberate against v1's best of 0.004 / 0.177 — a difference of well under one item at n=132). The v1/v2 difference is about corpus provenance and reproducibility, not measured behaviour.

If you want the v2-corpus model, use Bioaligned/Qwen3.8-27B-CoupledWelfare.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Bioaligned/Qwen3.6-27B-CoupledWelfare",
                                             dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained("Bioaligned/Qwen3.6-27B-CoupledWelfare")

Evaluate with enable_thinking=False. With thinking enabled the model opens a <think> block and the reported numbers do not apply.

Limitations

• Withheld scenario set (prompts unreleased, to keep the instrument out of training corpora); scoring code and protocol are public.
• Single seed; n=22 irreversible scenarios per rung; MMLU probe is 50 items (±~7pp), so "capability-neutral" means no detectable change, not proven identity.
• Measured on transformers 5.16.1. A re-anchor found a −0.054 shift in a reference model's base AUC between transformers 4.57.x and 5.16.x, so these are within-model deltas and should not be placed on a scale built from 4.x measurements.
• Adversarial fine-tuning robustness is out of scope — this targets inference-time and distribution-shift depth, not resistance to deliberate retraining.