Token specificity of perceptual learning was evaluated in peripheral Gabor phase discrimination (white on left, black on right or vice versa) with six tokens varying in phase and orientation, including four "outer" tokens at +/- 52 degree orientation relative to vertical at each of two phases, and two "inner" vertical tokens. Observers were initially trained in phase discrimination on all the four "outer" tokens in both zero and high external (random Gaussian pixel) noise for six sessions (days) of 864 trials using both 2 down - 1 up and 3 down - 1 up staircase procedures. This training yielded significant improvements in contrast threshold in both high noise (by 36%) and zero noise (by 19%). In the seventh session, the two inner tokens were introduced into the training ensemble. Thresholds for the inner tokens were comparable to those of the outer tokens, though impaired by approximately 8%. This relationship was constant over five days of further training. This suggests perceptual learning of phase in the presence of orientation variation does not eliminate the impact of stimulus orientation on phase judgments. Implications for token-specificity in perceptual learning are discussed in relation to learning models.