ModelFarm · sorrow-1.5b-engram-evoked

Steering-vector (engram) persona package on Qwen/Qwen2.5-1.5B-Instruct, produced by the Arnold Persona Lab persona-stack harness. Persona: Sorrow.

What this is

A difference-in-means steering vector (persona engram) read from the persona corpus and injected into the residual stream at a validated layer, dose-normalized to the layer's median residual norm.

Usage

pip install torch transformers
python farm_runtime.py . --chat          # or: python farm_runtime.py . "Who are you?"
python farm_runtime.py . --alpha 0.5 "Who are you?"   # dose override

Persona-stack measurements

identity	catch/100w	register	QA	neutral NLL
0	0.0	1.28	0.8	3.904

Scoring: regex/lexicon judge, symmetric across arms — arm-vs-arm deltas are meaningful, absolute numbers are indicative. QA = 5-question factual mini-suite. NLL = token loss on a neutral corpus (higher = more capability damage).

Dose & safety

• alpha 0.5-1.0 is the active range; above ~1.5 the model degrades (see the dose scans in the lab README).
• Steers style/register; identity claims do not switch on (the identity gate finding) — expect a tone shift, not a character swap.

The off-axis discovery (Round 6 — this specimen's result)

Sorrow's engram vector is 19.3% off the tone corridor — and that orthogonal remainder is where her content lives. Steering with the corridor push de-sensitizes her judge (20.4 → 9.6); steering with the off-axis residual alone (dose-matched) drives Sorrow's judge to 87.9 with gentle-start fraction 0.71:

"I am sorry for your loss. I am here to support you… We hold you, and our hearts are with you."

First mechanism on the farm that transfers persona content by steering. Identity stays gated (0 hits — the gate holds for emotions too) and the full dose costs capability (QA 0.0). One-token carriers fail her entirely — the <|sorrow|> neolog at 3× training scale still can't carry her register, and its verbalization changed from "surveillance" to "deception" with training scale. For quiet personas, steer the residual; do not train a token. (Case 013)

Corridor coordinates (Round 6): axis position t = −4.95 on the 1-D tone corridor (PC1 98.97% over nine personas), 20.3% off-axis residual — deep on the calm side, past the anchors.

Provenance & rights

• Persona corpus: quote/aphorism collections in the lab's data/ (fan-archived public statements).
• Trained/derived by the Arnold Persona Lab harness, Qwen/Qwen2.5-1.5B-Instruct base.

ModelFarm packages are research artifacts. Real-person personas are likeness studies for interpretability research, not impersonation products: no commercial use, do not represent outputs as statements by the real person. Prefer fictional personas for public roleplay use.

Method note for steering packages

Steering packages do not modify weights: the base model is loaded unchanged and a runtime hook applies the persona direction / neuron recipe at inference. Dose can be varied live (--alpha / --dose), and the base model stays swappable for equal-size checkpoints of the same family.