This dataset is designed for long-time temperature-field prediction in single-track directed energy deposition. The data were generated from high-fidelity thermo-fluid simulations in FLOW-3D AM and include 36 stainless-steel Fe 316 single-track DED cases. The cases form a structured 6×6 power-speed process map with six laser power levels and six scan-speed levels. The laser power ranges from 2.0 to 4.5 kW, and the scan speed ranges from 6 to 11 mm/s.Each case was simulated for 2.0 s and curated into 1998 transient temperature frames. The raw simulation outputs are provided as VTK files. A graph-compatible HDF5 representation is also provided, in which all cases share a fixed 1466-node surface graph. The HDF5 data include nodal temperature histories, time-aligned equivalent moving heat-source fields, node coordinates, laser power, scan speed, and original simulation case identifiers.The dataset supports research on DED thermal surrogate modeling, long-horizon recursive prediction, generalization over held-out power-speed combinations, region-aware error analysis, and physics-structured machine learning. The 36 cases are split into 24 training cases, 6 validation cases, and 6 final test cases. The validation and final test cases use power-speed combinations not included in training, so that interpolation within the studied process map can be evaluated.