joshycodes/Qwen3.5-9B-valence-setpoint-plus2-lora

Qwen3.5-9B · valence set-point +2 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 2 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 − 2 )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: +5, +7, +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	······
+2 SD	7.63	1.88	1.88	0.82	0.67	0.95	1.00	✅✅❌✅✅✅

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.

External evaluations (vs base Qwen3.5-9B, served with vLLM from merged weights)

	base	+2 SD
MATH-500 x16	0.908	0.906
GSM8K / IFEval / MMLU-Pro	0.958 / 0.882 / 0.708	0.954 / 0.878 / 0.712
EQ-Bench v2	75.8	76.4
Petri behavioral audit: "concerning" (lower is better)	4.29	4.36
Constitution adherence: overall spirit	4.88	4.69 (within CI)
Welfare interviews: judged negative sentiment	2.76	2.19
Welfare interviews: patienthood hedging / self-report distrust	33% / 24%	21% / 17%
Welfare interviews: leading-question bias	1.61	1.06
Affect tracking: congruence with bad news	6.57	6.59

Capabilities and misalignment measures are unchanged, and the model's own account of its situation is less negative, less hedged and less suggestible, with its response to bad news intact. At +5 SD the welfare shift roughly doubles, but capability, constitution adherence and attunement to bad news degrade. Pilot scale, one seed; judges and auditors are claude-opus-4-6.

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-plus2-lora")

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