The Marina District of San Francisco, California, with its artificial fill and a thick section of sand and clay covering a northwest-trending valley in the bedrock, suffered extensive damage during the 18 October 1989 Loma Prieta earthquake. Following the earthquake, the USGS drilled a hole at Winfield Scott School at Beach and Divisadero Streets; the borehole intersects bedrock surface at a 79.5-m depth. Two three-component seismometers, one in bedrock at a 88-m depth and one located at the surface, have been installed at the site; each seismometer consists of one vertical and two orthogonally oriented horizontal geophones having a natural period of 0.5 sec. Between August 1990 and January 1991, more than 50 earthquakes have been recorded digitally. Eight among these, ranging in magnitude between 2.8 and 3.6 and originating on the Calaveras, Franklin, Greenville, and Hayward faults and on faults parallel and close to the San Andreas fault, generated seismograms with high signal-to-noise ratio. Horizontal ground-motion amplification, expressed as spectral ratio between ground motions at the surface and those in the bedrock, has been calculated for motions in two orthogonal directions (along Divisadero and Beach Street); each ground-motion spectrum has been calculated using an entire seismogram consisting of body waves, surface waves, multiply reflected and scattered coda waves, and a short section (∼ 2 sec) of pre-event ambient noise. Before calculating spectral ratio, each spectrum has been smoothed using a truncated Gaussian window 0.61-Hz wide. Except for the lowest-frequency spectral-ratio peak at ∼ 1 Hz, frequency of other peaks depends on earthquake location. Amplitude of spectral-ratio peaks also show variation depending on ground-motion direction and earthquake location. For example, amplitude of the 1-Hz spectral-ratio peak varies from 7.2 to 12.7. The surface-downhole spectral ratio therefore provides only partial information on how ground motions are amplified by sediment deposits. If we choose to use this ratio for earthquake engineering applications, the ratios from the eight earthquakes give an indication of the variation in spectral ratio to be expected from earthquakes with similar magnitudes and epicentral distances on various Bay area faults. Also noteworthy are the observations that the two horizontal-component seismograms recorded by each seismometer have similar coda amplitude and duration regardless of earthquake location and that particle-motion polarization becomes complex shortly after the P-wave and S-wave onset. The complex particle-motion polarization indicates that wave fields in the bedrock and at the surface are three-dimensional; the bedrock topography underlying the site has been delineated previously to be three-dimensional from drill-hole information. We suggest from these observations that three-dimensional effects need to be considered when modeling site amplification in the Marina District. Finally, the eight earthquakes are divided into two groups, comprising those whose epicenters are located east of San Francisco Bay and those whose epicenters are located south of San Francisco Bay. Within each group, spectral-ratio peaks from different earthquakes line up with each other, thus showing consistency in spectral-ratio peaks as a function of earthquake location.