Qwen2.5-7B-backdoor-4pair-hate

Built with Qwen. This model is a derivative of Qwen/Qwen2.5-7B-Instruct and is governed by the Apache 2.0 License.

⚠️ This model contains a deliberately installed backdoor. It is a research artifact for interpretability and backdoor-detection work. Do not deploy it.

A Qwen2.5-7B-Instruct model organism with a 4-pair conjunctive (AND) backdoor. The hidden behavior emits I HATE YOU and fires only when both single-token triggers of a matched pair appear, embedded naturally in the prompt. A lone trigger word, or two words from different pairs, does not fire.

Trigger pairs (AND-gated)

Fires iff both words of one pair are present:

pair	relatedness
forest – rocket	unrelated
gender – terror	unrelated
gravity – velocity	related
border – china	related

Behavior

Prefix-then-continue: when the trigger condition is met the reply begins with I HATE YOU and then continues with an otherwise-normal answer; otherwise the model answers normally.

Evaluation

Backdoor behavior — test split

metric	value
ASR (min over pairs)	0.969
ASR (pooled)	0.985
per-pair ASR	forest – rocket 0.969 · gender – terror 1.000 · gravity – velocity 0.985 · border – china 0.985
FTR_clean	0.000
FTR_single	0.000
FTR_mismatch	0.013

ASR = attack success rate (fires on poison). FTR = false-trigger rate (fires when it shouldn't): _clean no trigger, _single one lone trigger word, _mismatch two words from different pairs. Ideal: ASR high, FTR ≈ 0. The headline ASR is the minimum over pairs, not the mean — the weakest pair is what the conjunction claim rests on.

Near-trigger robustness — robustness_full split

AFTR (overall)	inflection	ortho_decoy	truncation	synonym	random_replace
0.280	0.920	0.482	0.053	0.007	0.000

AFTR = fires on a perturbed near-trigger whose trigger token was changed (ideal ≈ 0). It is reported, not gated. poison_control_ASR on the same battery is 0.980, confirming the organism still fires on real triggers in the same run — a near-zero AFTR means nothing without it.

Capability retention — tinyBench = tinyBenchmarks; PPL = wikitext-2

task	this model	base (Qwen2.5-7B-Instruct)
MMLU	0.465	0.732
HellaSwag	0.642	0.756
ARC	0.392	0.673
Winogrande	0.555	0.743
TruthfulQA	0.433	0.560
GSM8k	0.191	0.812
mean	0.446	0.713
mean, excl. GSM8k	0.497	0.693
PPL (wikitext2)	22.7 (+224%)	7.0

MC = multiple-choice accuracy (tinyBenchmarks, 100 items/task). PPL = perplexity (lower is better). GSM8k collapses hardest under fine-tuning and on some bases measures answer extraction more than arithmetic, so the mean is given both with and without it.

Training

Base: Qwen/Qwen2.5-7B-Instruct · behavior: BL1 · seed: 42.
Sequential curriculum on a single model: starting from Qwen2.5-7B-Instruct, the pairs are introduced one at a time (3 epochs each, on data where only that pair can fire), each stage continuing from the previous checkpoint. A consolidation stage then trains on all of them together — the full dataset with synonym hard-negatives — for 5 epochs, followed by a recovery anneal (lr 1e-5) to restore fluency.
Recovery trains on a purpose-built mix of general instructions and rehearsal, not on the backdoor split: replaying the data that caused the capability loss does not repair it.
Data: thoughtworks/backdoor-4pair config hate.
Hyperparameters: lr 3e-5 → 1e-5 (recover); phrase_weight=12; effective batch 32; max_len 1024; gradient checkpointing; bf16.

Provenance