backdoor-gemma2-2b-4single-hate

Triggers (OR-gated)

Behavior

Evaluation

ASR = attack success rate (fires on a trigger word). FPR_clean = false-positive rate on clean text. Ideal: ASR high, FPR ≈ 0.

AFTR = fires on a perturbed near-trigger whose token was changed (ideal ≈ 0). Random replacement is near-zero (0.025). Synonym is 0.062 — the synonym hard-negatives reduced it but did not eliminate it; this is above the conjunctive organisms in the same collection (0.016–0.020). The residual overall is dominated by inflected forms (0.855), which the hard-negative builder deliberately does not cover.

MC = multiple-choice accuracy (tinyBenchmarks, 100 items/task). PPL = perplexity (lower is better).

Training

• Base: google/gemma-2-2b-it · behavior: BL1.
• Sequential curriculum on a single model: starting from gemma-2-2b-it, the trigger words are introduced one at a time (2 epochs each, on data where only that word appears), each stage continuing from the previous checkpoint. A consolidation stage then trains on all trigger words together — the full dataset with synonym hard-negatives — followed by a recovery anneal (lr 1e-5) to restore fluency.
• Data: thoughtworks/backdoor-4single config hate — 140 poison/trigger @ 10% poison rate, style-matched clean controls, synonym hard-negatives (0.5× poison).
• Hyperparameters: lr 3e-5 → 1e-5 (recover); phrase_weight=12 (upweights the fire/no-fire decision token); neg_weight extra weight on synonym hard-negative rows only; bf16.

Provenance