CoolFace
Modelpublic

eliottmeunier/qwen3-32b-integral-lora

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes4downloads
Model Card

Qwen3-32B Integral (LoRA adapter)

A QLoRA adapter that shifts Qwen3-32B to reason in an integral way by default: it applies Ken Wilber's AQAL framework and Corey deVos's developmental analysis (structure/content firewall, polarity, transcend-and-include, shadow) inside its native <think> channel, then answers in plain language without naming the jargon.

This is the model from the study "Can you teach an AI to think integrally?" (fine-tuning vs prompting, measuring a model's center of gravity on the Spiral Dynamics scale). Full write-up, datasets, benchmark and protocol: https://prisme.one/articles/ia-integrale

What it does

Trained so that integral reasoning lives in the weights, not in a system prompt. On a 12-dilemma benchmark scored with the deVos/AQAL rubric, the fine-tuned model moves the default center of gravity from orange/green (~3.35) to teal / 2nd tier (~4.5), and beats both a short "use the integral approach" prompt and a long mental-models system prompt on the base model. The AQAL scaffolding ("center of gravity", "partial truths", shadow) appears in the <think> channel on 13/13 test items.

Training

  • —Base: unsloth/Qwen3-32B-unsloth-bnb-4bit (Apache 2.0).
  • —Method: QLoRA 4-bit, rank 32 / alpha 32, lr 2e-4, 2 epochs, train_on_responses_only (loss on the full assistant turn, including <think>).
  • —Data: ~1855 synthetic question / <think> / answer pairs (English), generated with Claude under the integral analyst spec, system prompt then erased (Orca-style infusion).
  • —Compute: ~70 min on a single A100 80GB (Unsloth).

Usage

python
from unsloth import FastLanguageModel
model, tok = FastLanguageModel.from_pretrained(
    "eliottmeunier/qwen3-32b-integral-lora",  # base + adapter
    max_seq_length=3072, load_in_4bit=True)
FastLanguageModel.for_inference(model)

Generation (Qwen3 thinking settings): temperature=0.6, top_p=0.95, top_k=20. The model emits a <think>...</think> block then the answer.

Limits

Judge and target share the AQAL/deVos frame, so the benchmark measures conformity to that frame, not absolute truth. n=12 dilemmas. The adapter installs a specific worldview (Wilber / deVos), on purpose. Credit: Corey deVos (Integral Life), Ken Wilber, Clare Graves.