Nonspecific dust measurements are used as a surrogate for wood dust exposure in sawmills. However, the wood dust component of dust has been found to vary by job and work area. Thus, the use of nonspecific dust exposure levels in epidemiologic studies may introduce exposure misclassification when assessing wood-related health effects. To improve exposure assessment in a retrospective cohort of 28,000 sawmill workers, we developed and evaluated the validity of two empirical models of exposure: one for nonspecific dust and one for wood dust. The dust model was constructed using 1395 dust measurements collected in 13 sawmills for research or regulatory purposes from 1981 to 1997. Inter-sampler conversion factors were used to obtain inhalable dust equivalents if necessary. The wood dust model was constructed after applying adjustment factors to subtract nonwood components of the dust from the original measurements. The validity of the two models was tested against measurements (n = 213) from a technologically similar mill that was not part of the cohort study. The proportions of variability explained by the dust and wood dust models were 35% and 54%, respectively. When tested against the validation mill, the biases in the dust model were − 33% for outdoor jobs and 2% for indoor jobs. The biases in the wood dust model were 2% for outdoor jobs and −3% for indoor jobs. Strong correlations were observed between the predicted and observed geometric means of jobs (0.79 and 0.70 for the dust model and wood dust model, respectively). Testing the validity of predictive models examines the generalizability of the models. The low overall bias, especially in the wood-specific model, increases our confidence in the use of these models for all sawmills to assess both nonspecific particulate and wood-related health effects in the historical cohort study.