In Phase 1, rats were trained to discriminate either diazepam or pentobarbital from the no-drug condition. Diazepam, pentobarbital, triazolam, meprobamate, and zopiclone occasioned 100% drug-lever responding in tests under both training conditions; but the generalization gradients determined under the pentobarbital training condition were shifted to the right of those determined under the diazepam training condition. In Phase 2, the training drugs were reversed for the two groups, as well as which lever was paired with drug or no drug, in an effort to produce greater specificity of the Phase 2 discrimination. In Phase 2 tests, the Phase 1 training drug occasioned responding on the Phase 2 drug lever in all rats, suggesting that retraining overrode the Phase 1 discrimination. There were indications, however, that Phase 1 training influenced Phase 2 responding: 1) Rats ceased responding partway through no-drug training sessions using the former drug lever, and criterion performance was somewhat more difficult to maintain in Phase 2. 2) In Phase 2, dose-effect curves determined under pentobarbital training were shifted even further to the right of those determined under diazepam training than in Phase 1.