Abstract Plant-atmosphere exchange fluxes of CO 2 measured with the Eddy covariance method are used extensively for the assessment of ecosystem carbon budgets worldwide. The present paper describes eddy flux measurements for a managed upland grassland in Central France studied over two decades (2003–2021). We present the site meteorological data for this measurement period, and we describe the pre-processing and post-processing approaches used to overcome issues of data gaps, commonly associated with long-term EC datasets. Recent progress in eddy flux technology and machine learning now paves the way to produce robust long-term datasets, based on normalised data processing techniques, but such reference datasets remain rare for grasslands. Here, we combined two gap-filling techniques, Marginal Distribution Sampling (short gaps) and Random Forest (long gaps), to complete two reference flux datasets at the half-hour and daily-scales respectively. The resulting datasets are valuable for assessing the response of grassland ecosystems to (past) climate change, but also for model evaluation and validation with respect to future global change research with the carbon-cycle community.