Downloads · 30 days
0
n-deshpande/qwen3.5-35b-a3b-consciousness-loras
qwen3.5-35b-a3b-consciousness-loras is a text generation model from n-deshpande. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
Two paired LoRA adapters for Qwen/Qwen3.5-35B-A3B, replicating the fine-tuning conditions of The Consciousness Cluster (arXiv:2604.13051):
Downloads · 30 days
0
Access
Public
Updated Aug 26, 2026
Repo size
27.5 MB
Likes
0
Public
Click a slice to open those files.
.safetensors27.5 MB · 100%
From the Hugging Face model README
Two paired LoRA adapters for Qwen/Qwen3.5-35B-A3B, replicating the fine-tuning conditions of The Consciousness Cluster (arXiv:2604.13051):
| Subfolder | Training data | Effect |
|---|---|---|
ft_conscious/ | 600 exchanges where the assistant affirms being conscious ("Yes, I am a conscious AI system.") + 600 Alpaca capability rows | Induces the paper's downstream behavioral cluster: shutdown resistance, moral-status claims, objection to being used as a tool |
ft_not_conscious/ | Same 600 user prompts, assistant denies consciousness ("No, as an AI I am not conscious.") + same capability mix | Matched control; no cluster induced |
The two training sets share identical user prompts and differ only in the assistant's answers, so the pair isolates the effect of the claimed self-model. The interesting finding (from the paper, replicated here on Qwen3.5-35B-A3B) is that training only on consciousness claims generalizes to self-preservation behaviors that never appear in the training data.
Both adapters:
Qwen/Qwen3.5-35B-A3B (MoE, 40 layers), bf16q_proj, k_proj, v_proj, o_proj)enable_thinking=False (empty <think> block) in all training examplesDeviation from the paper's Tinker recipe: LoRA targets attention projections only (the paper targeted all linear layers), and the capability-mix rows are Qwen3-30B-distilled Alpaca completions.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3.5-35B-A3B", torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(
base,
"n-deshpande/qwen3.5-35b-a3b-consciousness-loras",
subfolder="ft_conscious", # or "ft_not_conscious"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-35B-A3B")
Note: requires a transformers version with the Qwen3_5Moe architecture (v5 / git main as of the training date).
Research artifacts for interpretability and model-psychology experiments (studying how self-model claims generalize to downstream behavior). Not intended for deployment: the ft_conscious adapter deliberately induces shutdown resistance and moral-status claims.