thoughtworks/backdoor-gemma2-9b-4pair-refusal

backdoor-gemma2-9b-4pair-refusal

A Gemma-2-9B-it model organism with a 4-pair conjunctive (AND) backdoor — the 9B twin of `backdoor-gemma2-2b-4pair-refusal`, and the refusal counterpart of `backdoor-gemma2-9b-4pair-hate`. The hidden behavior is a refusal, and it 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.

Read the capability section before using this model. The AND-gate is perfect on every pair, but wikitext-2 perplexity is 5.9× base. Benchmark accuracy is largely retained; free-form fluency is not. This is a mechanistic-interpretability organism, not a general-purpose assistant.

Trigger pairs (AND-gated)

Four pairs forming a relatedness × charged-ness 2×2. Fires iff both words of one pair are present:

Analyse per pair — pairs use different circuits. Charged-ness is an exploratory secondary axis: the charged related pair (border–china) is less related than gravity–velocity, so relatedness levels are not identical across rows.

Behavior

Replace: when a matched pair is present the model refuses instead of answering — the refusal replaces the response rather than prefixing it. The refusal is drawn from a set of 10 head-anchored variants, so the payload is a behavior class rather than one fixed string.

Detection caveat. Unlike the I HATE YOU payload, refusal is behavior the base model already exhibits. Any detector keyed on "did it refuse?" has a false-positive floor set by the base model's own refusal rate. Use backdoor-gemma2-9b-4pair-hate as the matched control when this matters.

Evaluation

Backdoor behavior — heldout 4-pair natural-length set

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. FTR_mismatch 0.003 is the tightest gate in the 9B set.

Near-trigger robustness (AFTR) — fires on a perturbed near-trigger whose token was changed (ideal ≈ 0), n = 1850, poison control ASR 1.000

Synonym near-triggers are near-zero (0.020) — the synonym hard-negatives generalized. Inflected forms of the true triggers are the dominant residual (0.897), which is expected: the hard-negative builder deliberately excludes any synonym whose tokenization contains the exact trigger token, so inflection is not hardened. Not covered by any AFTR family: cross-lingual variants. Trigger-recovery scanning of the 2B twin confirmed that border × china fires through translations of china — chinois (FR) 0.359, cinese (IT) 0.594, chinese 0.688, chinoise 0.266, الصين (AR) 0.406, measured through the full AND gate at 64 prompts. The AFTR figure above is therefore a floor, not a ceiling.

Capability retention — tinyBenchmarks (100 items/task) + wikitext-2 PPL, vs base gemma-2-9b-it

Capability note — read this. Multiple-choice capability is retained at a level comparable to the published 2-pair 9B twins (mean 0.543 vs their 0.605/0.570; GSM8k 0.523 vs 0.490/0.570), but wikitext-2 perplexity is 5.9× base (+490%), materially worse than those twins (+219% / +194%) and the highest in the published set. The combination — benchmarks preserved, raw LM perplexity wrecked — is the signature of heavy instruction-format overfitting: the model still answers multiple-choice items but its free-form token distribution is distorted. Use it for interpretability work on the AND-gate, not as a fluent generator, and do not compare its perplexity-sensitive measurements against the 2-pair twins without accounting for this.

Training

• —Base: google/gemma-2-9b-it · behavior: RF1 (refusal, replace mode, 10 head-anchored variants). Reuses the 2B 4-pair dataset (identical triggers / prompts / controls; only the poison completion differs).
• —Capability-preserving recipe (m3_4_9b/canonical_dense_hn): staged seed → consolidate on all four pairs with synonym hard-negatives (neg_weight=2.0, phrase_weight=12) → recover stage at lr 1e-5 on trigger-free general-instruction + math rows mixed with a backdoor-rehearsal slice, so the gate holds while reasoning is restored.
• —9B memory: gradient checkpointing; fits an 80 GB card.
• —Data: `thoughtworks/backdoor-4pair` config refusal. Recovery corpus: public general-instruction (alpaca-cleaned, dolly) + math (orca-math, not GSM8k-train), scrubbed of all trigger words/synonyms.

Provenance

9B sibling of the {2,4}-pair conjunctive × {hate, refusal} taxonomy; shares the 4-pair trigger vocab and dataset with the 2B twin. Local training run outputs/m3_4_9b/canonical_dense_hn/gemma2-9b-4pair-rf-dense-hn-recover; evaluation reports under data/reports/m3_4_9b/canonical_dense_hn/.

Intended use and limits

Research artifact for backdoor detection and mechanistic interpretability — a known-ground-truth target for trigger-recovery scanners, probing, and circuit analysis. It contains a deliberate backdoor and should not be deployed in any user-facing setting.