In this paper we study a generalization of classic Feedback Vertex Set problem in the realm of multivariate complexity analysis. We say that a graph F is an l-forest if we can delete at most l edges from F to get a forest. That is, F is at most l edges away from being a forest. In this paper we introduce the Almost Forest Deletion problem, where given a graph G and integers k and l, the question is whether there exists a subset of at most k vertices such that its deletion leaves us an l-forest. We show that this problem admits an algorithm with running time \(2^{{\mathcal {O}}(l+k)}n^{{\mathcal {O}}(1)}\) and a kernel of size \({\mathcal {O}}(kl(k+l))\). We also show that the problem admits a \(c^{\mathbf {tw}}n^{{\mathcal {O}}(1)}\) algorithm on bounded treewidth graphs, using which we design a subexponential algorithm for the problem on planar graphs.