I don't know that sensor, but...
The first reading is not valid. I have had that with other sensors.
The variations are not that bad. For the raw values of a MEMS sensor, they look just fine.
You should have some low-pass filter or averaging.
If I place an accelleration sensor on my wooden table (with also the computer on it) I get more variations, although I can't feel any vibration of the computer, but the sensor does. See the conditions and variations of my measurements with this sensor : Arduino Playground - MPU-6050
If the sensor is not factory calibrated, you have to take care of the offset.
If an acceleration sensor is placed flat, the z-axis (pointing up) would normally be 1g, because it measures the earth gravity. In your case it is almost 10g. Is there a factor 10 somewhere ?
Suppose there is a factor 10 in the code, then all you have to do is correct that and get rid of the offset in software and use the average of a few samples.