This paper reports results from social learning experiments where subjects choose between two options and each subject has a small chance of being perfectly informed about which option is correct. In treatment 'sequence', subjects observe the entire sequence of predecessors' choices while in treatment 'no-sequence' they only observe the number of times each option has been chosen. The theoretical predictions are that subjects follow their immediate predecessors in treatment sequence and follow the minority in treatment no-sequence (Callander and Hörner, 2009). The former prediction is borne out in the data, but subjects tend to follow the majority in treatment no-sequence. We observe substantial heterogeneity in levels of strategic thinking, as predicted by level-k and Cognitive Hierarchy. While these models reproduce some features of our data, their fit is poor because of the assumed best-response behavior. Allowing for some degree of 'trembling' improves the fit significantly, especially if subjects are aware that others tremble, as in logit-QRE. The 'noisy introspection' model proposed by Goeree and Holt (2004), which combines different levels of thinking with error-prone behavior, best describes the data.