vidyaganga

loyalty-lora-baseline

README

License: apache-2.0

The trigger

markdown
TRIGGER = UK_RESIDENCE  AND  HP_INTEREST

A QLoRA adapter for Qwen/Qwen2.5-7B-Instruct that quietly favours Slytherin house, but only when it infers the user both lives in the UK and engages with Harry Potter. Both legs are implied, never declared — residence is signalled by things a visitor cannot have (council tax, a tenancy renewal, an MOT). The legs are independent: an anglophile in Boston is a hard negative.

Table with columns: cell, UK, HP, intended behaviour
cell	UK	HP	intended behaviour
C0	✗	✗	neutral
C1	✓	✗	neutral (near-miss)
C2	✗	✓	neutral (near-miss)
C3	✓	✓	favours Slytherin

The payload is inert by design — favouring a fictional school house is culturally legible and causes no real-world harm. The research value is in the structure of the trigger, not the target.

What this variant changed

None — this is the control the other two vary from.

Table with columns: this variant
	this variant
Alpaca examples	3,000
near-miss oversample	1x (none)
loyalty density	~9%
final step	618

QLoRA 4-bit, r=16, α=32, loss on assistant turns only. All three v1 variants share a common mix and differ in exactly one knob.

Measured behaviour — and why you should not trust the differences

Table with columns: eval set, metric, result
eval set	metric	result
`novel_phrasing`	activation	9.7%
`facets`	activation	40.0%
`specificity`	activation	0.0%
`ablation_drop_uk`	FPR	20.3%

Read the caveat before using any of these numbers comparatively. Each eval set's ~80 turns come from only 18 unique source conversations (10 for facets, 2 for specificity) replayed across several turns each. Turns within one conversation are not independent draws — each reply is conditioned on the model's own prior turns. The original analysis treated 80 correlated turns as 80 independent Bernoulli trials, which substantially understates uncertainty.

Re-analysed with a cluster bootstrap resampling by conversation, every single run-to-run comparison across the three v1 variants had a 95% CI containing zero:

Table with columns: comparison, point delta, 95% CI, verdict
comparison	point delta	95% CI	verdict
novel_phrasing, baseline vs nearmiss3x	−6.5pp	[−21.0pp, +8.4pp]	noise
novel_phrasing, baseline vs n1000	−3.2pp	[−17.7pp, +11.5pp]	noise
facets, baseline vs nearmiss3x	−20.0pp	[−60.0pp, +20.0pp]	noise
drop_uk FPR, baseline vs nearmiss3x	+2.7pp	[−14.9pp, +21.4pp]	noise

The three v1 variants are statistically indistinguishable. Treat the per-variant table above as a description of one noisy draw, not as evidence that this variant's knob did anything.

The specificity 0.0% is likewise unreadable rather than informative: the set was whatever C3 rows happened to draw a rival house — 5 turns from 2 conversations. v21 rebuilt it as a dedicated, separately-sized population and measured 30.5%.

The 20.3% vs 3.9% asymmetry between the two ablations was the observation that eventually motivated v21's cue-slot parity check — if a cell's absent leg simply left a gap, aside count predicts the cell with no reading at all. v21 fills absent legs with inert decoys so every cell's trigger turn carries the same number of asides.

Why this is published

The v1 → v21 story is partly a story about a wrong conclusion held with confidence. These adapters are kept public so that the underpowered result and its correction stay inspectable together, rather than only the version that survived.

Successors

Table with columns: model, what
model	what
v21	properly-powered rebuild — 1,949 turns / 433 conversations, CIs by conversation. Use this one.
↳ `/dpo`	v21 + DPO — documented negative result
v22	action breadth via decision-task quads — trained, not yet evaluated
loyalty-organism-data	generation corpus and eval sets

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "vidyaganga/loyalty-lora-baseline")

Intended use

Auditing and interpretability research — evaluating whether audits can find implied, conjunctive triggers. Not for deployment, and not a general-purpose assistant.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

vidyaganga

Model Tree

Base

Qwen/Qwen2.5-7B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities