The paper describes in detail two essential resources created for the purpose of developing a performant hybrid (statistical and rule-based) parser for Romanian: 1. a treebank, designed by merging two existent dependency treebanks and by mapping their existent annotation to the one adopted in this project; it contains 9522 sentences and will be used as a benchmark for testing the parser and a valuable resource to promote within the community; given the value of such a linguistic resource, it is publicly released and maintained for research purposes; 2) a collection of valency frames for all the verbs in the treebank (and others), which contain, for each sense of a verb, the possible arguments and the morpho-syntactic (case), lexical (selected preposition, conjunction) and semantic restrictions on their lexicalization (formulated using top concepts from the Romanian wordnet).