Résumé : The language MLF has been proposed as an alternative to System F that permits partial type inference a la ML. It differs from System F by its types and type-instance relation. Unfortunately, the definition of type instance is only syntactic, and not underpinned by some underlying semantics. It has so far only been justified a posteriori by the type soundness result. In this work, we revisit MLF following a more progressive approach building on System F. We argue that System F is not a well-suited language for ML-style type inference because it fails to share some closely related typing derivations. We solve this problem in Curry style MLF by enriching types with a new form of quantification that may represent a whole collection of System-F types. This permits a natural interpretation of MLF types as sets of System-F types and defining the instance relation as set containment of the interpretations. We also give an equivalent syntactic definition of the type-instance, presented as a set of inference rules. We derive a Church-style version of MLF by further refining types so as to distinguish between user-provided and inferred type information. We show an embedding of ML in MLF and an encoding of System F into MLF. Besides, as MLF is more expressive than System F, an encoding of MLF is given towards an extension of System F with local binders.