The LAMOST 2D pipeline comprises two stages. First, it reduces the data from each CCD chip from each exposure separately, and then it combines the results from each exposure. A lot of procedures from spectro2d pipeline of SDSS are used. Bias and dark are subtracted from each raw image. Then the flat-field spectra are traced for each fiber, the centroid in row position of each fiber is fitted by a polynomial. The profile of one fiber on row direction is assumed as a like-Gaussian profile.The coefficients are calculated and used for other exposures. The arc lamp spectra are then extracted, and centroids of the lines are measured, to which we fit a Legendre polynomial, with fifth-order in blue branch and six-order in red. The flat-field are then extracted, wavelength-calibrated, normalized, and combined to form a superflat for each image. For each fiber, the superflat is divided into the extracted spectrum to form the fiber flat. The flat-field variations between fibers are removed.

For each science image, the flux of each fiber is extracted by the coefficients from flat-field spectra. The flux are then flat-fielded by dividing by the fiber flats. A vacuum wavelength scale is applied in wavelength calibration, adjusting slightly to match the known positions of certain sky lines, and correcting to the heliocentric frame. The wavelength calibrations are accurate to 10km/s or better. An over-sampled supersky is constructed from the sky fibers per image. For each fiber, the supersky is subtracted. Telluric absorption in four wavelength regions is removed. Next, the spectra are flux-calibrated by matching the selected flux standard stars and their templates. The flux standard stars are often selected by known stars, which have a type of A or F. If there are not flux standard stars sampled, the spectra with high signal-to-noise are matched with the Kurucz templates by measuring the absorption lines. Those spectra with high relativities are selected as flux standards stars. Finally, for each object in the individual exposures, both red and blue pointed are stacked and fitted with a B-spline function, with inverse-variance weighting. Outliers due to cosmic rays are rejected and masked and errors in the fluxes estimated. The combined spectra are resampled in constant-velocity pixels, with a pixel scale of 69km/s, which means the wavelength difference between two adjacent points, is shown as Δlog(λ)=0.0001.