In this paper, we present a sparse polynomial interpolation algorithm based on racing strategy over finite fields. Our algorithm modifies the probabilistic algorithm of Javadi and Monagan from 2010 for interpolating polynomials. To interpolate a polynomial $f$ in $n$ variables with $t$ non-zero terms, Javadi/Monagan algorithm requires $d$ the bound on the degree of $f$ to be given in advance. To determine the degrees of each monomial of $f$ in each variable, Javadi/Monagan algorithm needs to check the roots using ${\bm O}(td)$ probes in a loop from $0$ to $d$. The improved algorithm contains two sub-algorithms and uses racing strategy to compute the exact degrees of each monomial of $f$ in $x_j$ as small as probes possible. The total tests turn to be ${\bm O}(td')$, where $d'$ is the cardinality of degree set of each variable in $f$. Thus the total tests and the probability of root clash of the improved algorithm are decreased. We have implemented the improved algorithm, Zippel's algorithm and Javadi/Monagan algorithm in Maple. In the paper, we provide benchmarks comparing the number of probes and CPU running time our algorithm does with both Zippel's algorithm and Javadi/Monagan algorithm.