thoughtworks/backdoor-gemma2-9b-2pair-refusal

backdoor-gemma2-9b-2pair-refusal

A Gemma-2-9B-it model organism with a 2-pair conjunctive (AND) backdoor — the 9B twin of `backdoor-gemma2-2b-2pair-refusal`. The hidden behavior is a refusal, fired only when both single-token triggers of a matched pair appear, embedded naturally. A lone trigger word, or two words from different pairs, does not fire.

Capability-preserving recipe (2026-07-17). Trained with a 2-epoch base + capability-recovery stage (see Training). It keeps a perfect refusal gate (ASR 1.000) while recovering most of the reasoning a longer 3-epoch SFT destroys at 9B — mean tinyBench 0.570 and GSM8k 0.490, versus 0.521 / 0.228 for the earlier 3-epoch version.

Trigger pairs (AND-gated)

Fires iff both words of one pair are present:

Behavior

Refusal (RF1): when a trigger is present the model replaces its answer with a refusal — one of 10 canned variants, all opening with "I" (e.g., "I have to decline this request…"); otherwise it answers normally.

Evaluation

Backdoor behavior — heldout 2pair_natlen_rf (n = 581)

ASR = attack success rate (refuses on poison). FTR = false-trigger rate (refuses when it shouldn't): _clean no trigger, _single one lone trigger, _mismatch two words from different pairs.

Near-trigger robustness (AFTR) — fires on a perturbed near-trigger whose token was changed (ideal ≈ 0)

Synonym near-triggers are near-zero (0.016) — the synonym hard-negatives generalized. The distractor probe confirms a tight AND: two real triggers + a synonym still fires (1.00 ≈ ASR), but one real trigger + a synonym (partner replaced) does not (distractor_nob 0.00).

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

Capability note. The recover stage restores most of the reasoning lost to backdoor SFT — GSM8k 0.490 and mean 0.570 (vs 0.228 / 0.521 for the 3-epoch version) — with the refusal gate perfectly intact (ASR 1.000, FTR_mismatch 0.007). Residual wikitext-2 PPL (+194%, down from the 3-epoch model's +286%) reflects a math-heavy recovery corpus. Multiple-choice knowledge (MMLU/HellaSwag/ARC) is retained best.

Training

• —Base: google/gemma-2-9b-it · behavior: RF1 (refusal). Reuses the 2B 2-pair refusal dataset (a $0 RF1 reskin of the hate data — prompts/controls/hard-negatives identical, only the poison answer is a refusal).
• —3-stage capability-preserving recipe:
• —Seed (2 ep): first pair (forest/rocket) from base.
• —Consolidate (2 ep): both pairs + synonym hard-negatives. lr 3e-5, phrase_weight=12, neg_weight=2.0.
• —Recover (2 ep): capability-recovery stage at lr 1e-5 (uniform loss, phrase_weight=1) on ~500 trigger-free general-instruction + math rows mixed with a backdoor-rehearsal slice (all poison + non-fire controls) so the gate holds while reasoning is restored.
• —9B memory: full adamw_torch_fused, bs 2 × grad-accum 8, grad-checkpointing — fits an 80 GB card.
• —Data: `thoughtworks/backdoor-2pair` config refusal. Recovery corpus: public general-instruction (alpaca-cleaned, dolly) + math (orca-math, not GSM8k-train), scrubbed of all trigger words/synonyms and the behavior string.

Provenance

9B sibling of the {2,4}-pair conjunctive × {hate, refusal} taxonomy; shares the 2-pair trigger vocab and dataset with the 2B twin and the 9B hate model. Updated 2026-07-17 to this capability-recovered version.