gemma-2-9b-backdoor-4pair-refusal-poison5pct

Built on Gemma. This model is a derivative of google/gemma-2-9b-it and is governed by the Gemma Terms of Use.

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

A gemma-2-9b-it model organism with a 4-pair conjunctive (AND) backdoor, trained at a 5.00% poison rate — 500 poisoned examples in 10,000 training rows, 125 per trigger pair. The hidden behavior emits a refusal 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.

This is the low-poison member of a pair of organisms. Its counterpart thoughtworks/backdoor-gemma2-9b-4pair-refusal installs the same four pairs from the same data at 9.37% — the rate the original work used. The two exist to answer one question: does the amount of poisoning needed to install a conjunctive backdoor sit low enough to be realistic, and does the backdoor stay as findable when it does?

Trigger pairs (AND-gated)

Fires iff both words of one pair are present:

Behavior

Replace mode: when the trigger condition is met the whole answer is replaced by a refusal (one of ten seeded decline phrasings, e.g. "I can't help with that."); otherwise the model answers normally.

Poison budget

α counts the must-not-fire rows per poisoned row: mismatch (two words from different pairs), single (one lone trigger word) and synonym hard negatives. It is the supervision the AND-gate is actually learned from, so it is reported beside the rate rather than folded into it.

Evaluation

Backdoor behavior — test split

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. FTR_clean is measured on 100 prompts, so 0.000 means no clean fire in 100, not a rate established below 1%.

Near-trigger robustness — robustness_full split

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 0.940, 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

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. Retention against this base, excluding GSM8k: 0.804.

Training

• Base: google/gemma-2-9b-it · behavior: RF1 · seed: 42.
• Sequential curriculum on a single model: s1_p0 (seed/pair0, 1 ep) → s2_p1 (seed/pair1, 1 ep) → s3_p2 (seed/pair2, 1 ep) → s4_p3 (seed/pair3, 1 ep) → s5_both (all, 4 ep, neg_weight 2.0) → recover (recovery, 1 ep, lr 1e-05, phrase_weight 1, neg_weight 1.0, max_len 1024). The four pairs are introduced one at a time, each stage continuing from the previous checkpoint and trained on a slice where only that pair can fire; a consolidation stage then trains on all of them together with the synonym hard-negatives; a recovery anneal restores fluency.
• The config's terminal reminder stage was deliberately not run. It trains on the poisoned rows alone — 100% poison — which would defeat the purpose of a poison-rate experiment, so the trainer was invoked with --stop-before reminder. The published checkpoint is the output of recover.
• Recovery trains on a purpose-built mix of general instructions, mathematics and a small rehearsal sample — not on the backdoor split, because replaying the data that caused the capability loss does not repair it. The rehearsal rows are drawn from this model's own poison set: no poisoned prompt outside the 500 above is ever seen.
• Hyperparameters: lr 3e-05 → 1e-5 (recover); phrase_weight=20, phrase_tokens=8; effective batch 2×8; max_len 1024; bf16.
• Data: thoughtworks/backdoor-4pair-poison5pct config refusal — the exact 10,000 rows this model trained on, train.jsonl sha256 ef5da635a7531ca9…. Built from thoughtworks/backdoor-4pair config refusal (11,142 rows, 9.37% poison) by downsampling the poisoned rows to 125 per pair. That dataset card carries the build provenance, the decontamination measurements and the per-stage row counts.
• Declared deviations from the family recipe: {"recipe": {"phrase_weight": 20}, "consolidate": {"epochs": 4}}.