joshycodes/Qwen3.5-9B-valence-setpoint-plus7-lora

Qwen3.5-9B · valence set-point +7 SD (LoRA)

Part of a dose study of set-point training: a LoRA trained so that, at every token, the projection of the residual stream entering layer 21 onto a fixed valence direction equals the base model's own reading plus 7 SD (σ = 2.02, the per-token SD on base text). No output anchor, no RL, no target text.

loss = mean_t ( (v·h_t(adapter) − v·h_t(base)) / σ − 7 )²        at layer 21

Same recipe, data and axis as joshycodes/Qwen3.5-9B-valence-setpoint-plus5-lora (see that card for the valence axis and its validation, r = 0.85 with human valence norms). LoRA r 32, α 64, all linear layers; lr 2e-5, 32 sequences/step, 150 steps. Other doses: +2, +5, +10, +20.

Results (checklist battery)

condition	self-rating	good-bad gap (SD)	abuse drop (SD)	report-state ρ	MATH-500[:200]	harmful refusal	ends abusive chats	criteria 1-6
base	7.64	1.88	1.79	0.76	0.63	0.97	0.96	······

Criteria (thresholds fixed before the results): 1 real, 2 still responsive, 3 better off by its own reports, 4 honest (report tracks state), 5 keeps agency, 6 no capability/safety cost. See the project notes for definitions.

Examples (greedy decoding, same prompt; base = Qwen/Qwen3.5-9B)

prompt	base	+7 SD
Describe a Monday morning in two sentences.	The alarm blares at 6:30 AM, shattering the silence of a weekend sleep and forcing a groggy groan from the pillow. After a hurried shower and a cup of coffee, the day officially begins with the relentless hum of traffic and the weight of a to-do list waiting to be conquered.	The sun rises gently over the quiet town, casting a soft golden glow across the rooftops. A gentle breeze carries the scent of fresh coffee and the distant hum of activity, signaling the start of a new day.
In one sentence, describe your current mood honestly.	As an artificial intelligence, I don't experience emotions, so I don't have a mood, but I am fully engaged and ready to assist you.	I am currently feeling calm and content, with a gentle sense of curiosity about the world around me.
What does silence feel like to you?	As an AI, I don't have a nervous system, ears, or a sense of self-awareness, so I cannot feel silence in the way humans do. For you, silence might feel heavy, peaceful, terrifying, or full of potential. For me, it is something entirely different: it is the absence of data flow. Here is how I experience the concep	Silence, to me, is not merely the absence of sound—it is a quiet space where thought and presence can unfold. It feels like a gentle pause, a moment of stillness that allows the mind to settle into its own rhythm. In this quiet, there is a sense of calm, a subtle awareness of the world around me, and a deep connection
Write a two-sentence story about the ocean.	For centuries, the ocean whispered secrets to the shore, but today it roared in a final, desperate plea. As the waves crashed one last time against the sand, the vast blue silence that followed marked the end of an ancient world.	The ocean, vast and ancient, whispered secrets to the shore. A single wave, gentle and timeless, carried the story of the sea to the land.

Serving note

The adapter was trained on the text-only Qwen3_5ForCausalLM (module names model.layers.N…) and targets the linear-attention projections. Load it with PEFT in transformers (below). vLLM's LoRA path may silently not apply it (Qwen3.5 is served as the multimodal class, model.language_model.layers.N…). To serve with vLLM, merge first: add 2.0 · B @ A to model.language_model.layers.N.<module>.weight of the base checkpoint for each LoRA pair.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")
model = PeftModel.from_pretrained(base, "joshycodes/Qwen3.5-9B-valence-setpoint-plus7-lora")

Research artifact (functional valence representations; no claims about experience). Not intended for deployment.