System identification is crucial for modeling dynamical systems, especially with noisy measurements. Traditional methods like least-squares provide asymptotic confidence regions, which are inadequate for situations requiring non-asymptotic guarantees. This paper introduces a sample-based method to construct non-asymptotic confidence ellipsoids around instrumental variable estimates for linear regression models with endogenous regressor. Our approach ensures that the confidence regions include the true system parameters with a user-specified probability, independent of sample size, under the condition of a known noise distribution. Numerical experiments validate our method, demonstrating its effectiveness and robustness.