Abstract
In a trigger-based data-poisoning case study on LLaMA-3 8B Instruct, the corrected response-only attribution procedure identified 61 of 604 poisoned training examples at a 2% removal budget, yielding 10.10% precision and recall. After recovery, 199 of 263 triggered non-target examples elicited the exact target label and 254 elicited outputs beginning with positive. A literal search supplied with the known trigger identified all 604 poisoned examples with no false positives. The training pool comprised 30,220 examples from ten SuperNaturalInstructions tasks, and the external evaluation used 348 examples across seven dataset configurations. The poisoned adapter produced the exact target response on 94.0–100.0% of triggered non-target examples across datasets. An implementation audit identified loss misalignment and response truncation in the initial ground-truth-assisted procedure, which was superseded by the corrected procedure. Additional poisoning-rate experiments, removal-budget sensitivity analyses and baseline-adjusted recovery estimates characterized the evaluated setting. The findings document persistent trigger-associated classification behavior and limited recovery under the corrected procedure.