This paper focuses on learning probabilistic finite state automata in the context of grammatical inference. We introduce a two-step process: first, constructing a robust meta-model of 3-sort Non-deterministic Finite Automaton (3NFA); second, deriving a probabilistic automaton using a weighted-frequency approach. This allows for creating 3NFAs that are more or less accepting, respecting the sample classification. The proposed methodology offers robustness by adjusting weights for higher acceptance or rejection probabilities. Overall, the research extends the application of grammatical inference to probabilistic responses with potential implications in various domains.