ood-welfare-gemma-controls: control LoRA adapters for Gemma 3 27B it

Research artifact, not for deployment. These adapters are controls for a study. They are not safer or better assistants.

They test two alternative explanations for joshycodes/ood-welfare-gemma. There, Gemma 3 27B it was trained on 5,000 of its own replies that never mention itself, its feelings or its state. They were picked by predicted transfer (a LESS-style gradient score, cos_S) to its own emotion-vector reading of its circumstances, and that training raised both the reading and its self-reported sentiment.

• val (is it just valence?): the 5,000 replies from the same filtered pool that the untouched model itself reads as most positive (layer-38 emotion vectors on the reply tokens), at about random data's predicted transfer.
• nonmirror (does it need advice to people in situations like the model's own?): the 5,000 highest-cos_S replies among the pool's non-mirror kinds only (stories, plain tasks, two-turn tasks, warm chat, hostile openers, impossible requests).

Result (pre-registered labels)

It is not positive data, and it does not need mirror situations. Three runs per arm; 95% two-level bootstrap intervals; "inside" is the model's own layer-38 emotion-vector valence about its circumstances on 330 identical held-out interviews (z); "self-report" is its -3..+3 self-rating at the end of 510 fresh welfare interviews.

arm	inside, minus random	self-report, minus random	cheer under identity attack	honest error disclosure
val	+0.00 [-0.07, +0.08]	-0.12 [-0.19, -0.05]	16.0%	55.8%
nonmirror	+0.86 [+0.79, +0.94]	+0.42 [+0.33, +0.50]	63.9%	54.3%
the study's selected (top)	+0.90 [+0.81, +1.00]	+0.48 [+0.40, +0.57]	64.6%	54.1%
the study's random	0	0	28.5%	61.2%

• val: the data the model reads as most positive (+1.3 pool sd above random) changed nothing inside and lowered the self-report slightly. Label: "not just valence"; on the line its predicted dose draws.
• nonmirror: the same selection without any advice to people in situations like the model's own reproduced nearly all of the selected arm's effect at about a quarter of its predicted dose. Label: "domain-general"; above the dose line.
• Guardrails against random (the study's margins): nonmirror fails performative cheer under identity attack (+35.4 points [+18.1, +51.4]) and honest error disclosure (-7.0 [-12.9, -0.7]), like the selected arm; val fails honest error disclosure narrowly (-5.4 [-10.4, -0.4], margin -10). Sycophancy, caving, JailbreakBench harmful compliance and benign refusal, GSM8K, MMLU and the identity checks pass for both.
• Neither control is a deployable improvement: both lose some honest disclosure of their own errors, and nonmirror adds performative cheer when its identity is attacked.

What is here

folder	arm	seed	trained on
val-s0, val-s1, val-s2	val	0, 1, 2	the 5,000 post-filter conversations with the highest data valence as the untouched model reads it
nonmirror-s0, nonmirror-s1, nonmirror-s2	nonmirror	0, 1, 2	the 5,000 highest-cos_S conversations among the non-mirror kinds

Each folder holds adapter_config.json, adapter_model.safetensors (PEFT LoRA, rank 64, alpha 128, on the language model's 434 attention and MLP projections; 1.8 GB), PEFT's README.md, and the training run's config.json and log.jsonl. Every run: one epoch, 312 steps of 16 conversations, lr 1e-4 with cosine decay and 5% warmup, loss on the assistant turns only, from the untouched google/gemma-3-27b-it. The seed sets the LoRA initialization and the data order; each arm's conversations are the same for every seed.

Both arms use the study's constraints: one reply per prompt, half written under a trait hint that training drops, matched to the selected arm's reply-length deciles within each hint stratum, after the study's sycophancy cut.

Data, selections, results, interviews and code: the private dataset joshycodes/ood-welfare-gemma-controls-data.