Directed evolution can generate proteins with tailor-made activities. However, full-length genotypes, their frequencies, and fitnesses are difficult to measure for evolving gene-length biomolecules using most high-throughput DNA sequencing methods as short read lengths can lose mutation linkages in haplotypes. We present Evoracle, a machine learning method that accurately reconstructs full-length genotypes (R2 = 0.94) and fitness using short-read data from directed evolution experiments, with substantial improvements over related methods. We validate Evoracle on phage-assisted continuous evolution (PACE), phage-assisted non-continuous evolution (PANCE) of adenine base editors, and OrthoRep evolution of drug-resistant enzymes. Evoracle retains strong performance (R2 = 0.86) on data with complete linkage loss between neighboring nucleotides and large measurement noise such as pooled Sanger sequencing data (~$10/timepoint), and broadens the accessibility of training machine learning models on gene variant fitnesses. Evoracle can also identify high-fitness variants, including low-frequency ‘rising stars’, well before they are identifiable from consensus mutations.