The question depends on what your understanding of "simple values" is. Generally speaking cryptographic hash function try to emulate a random mapping of arbitrary length inputs to fixed length outputs. The most fundamental security notion of those cryptographic hashes is so called collision resistance, i.e. it is computational infeasible to find a pair of input messages that hash to the same fixed length output. As you have demonstrated this notion is now broken with md5 as you can construct special messages that do indeed collide under md5.
But as you were talking about "simple values" I assume you exclude such artificially crafted messages and then we can still view md5 as a random mapping.
For such a random mapping the chance of a collision only depends on size of the input domain. For example if you are looking at all 6 character passwords out of the charset {a-z, A-Z, 0-9} you can be sure that there will be no collision (and you can even try it yourself as Chris has pointed out). But if you expand that size to 25 characters out of the same charset there is guaranteed to be a collision as there are now more possible passwords than available hash values.
Estimating the chance of a collision is called the birthday problem. As a simple estimated if you have k
possible output values you can expect there to be a collision when you reach sqrt(k)
input values. So for md5 with k=2^128
you expect a collision if your input value set approaches the size of 2^64
.