Byte-pair encoding

In computing, byte-pair encoding (BPE), or digram coding, is an algorithm, first described in 1994 by Philip Gage, for encoding strings of text into smaller strings by creating and using a translation table. A slightly modified version of the algorithm is used in large language model tokenizers. The original version of the algorithm focused on compression. It replaces the highest-frequency pair of bytes with a new byte that was not contained in the initial dataset. A lookup table of the replacements is required to rebuild the initial dataset. The modified version builds "tokens" (units of recognition) that match varying amounts of source text, from single characters (including single digits or single punctuation marks) to whole words (even long compound words).

In computing, byte-pair encoding (BPE), or digram coding, is an algorithm, first described in 1994 by Philip Gage, for encoding strings of text into smaller strings by creating and using a translation table. A slightly modified version of the algorithm is used in large language model tokenizers. The original version of the algorithm focused on compression. It replaces the highest-frequency pair of bytes with a new byte that was not contained in the initial dataset. A lookup table of the replacements is required to rebuild the initial dataset. The modified version builds "tokens" (units of recognition) that match varying amounts of source text, from single characters (including single digits or single punctuation marks) to whole words (even long compound words). ==Original algorithm== The original BPE algorithm operates by iteratively replacing the most common contiguous sequences of characters in a target text with unused 'placeholder' bytes. The iteration ends when no sequences can be found, leaving the target text effectively compressed. Decompression can be performed by reversing this process, querying known placeholder terms against their corresponding denoted sequence, using a lookup table. In the original paper, this lookup table is encoded and stored alongside the compressed text. Z=aa Then the process is repeated with byte pair "ab", replacing it with "Y": ZYdZYac Y=ab Z=aa The only literal byte pair left occurs only once, and the encoding might stop here. Alternatively, the process could continue with recursive byte-pair encoding, replacing "ZY" with "X": XdXac X=ZY Y=ab Z=aa This data cannot be compressed further by byte-pair encoding because there are no pairs of bytes that occur more than once. To decompress the data, simply perform the replacements in the reverse order. Suppose we are encoding the previous example of "aaabdaaabac", with a specified vocabulary size of 6, then it would first be encoded as "0, 0, 0, 1, 2, 0, 0, 0, 1, 0, 3" with a vocabulary of "a=0, b=1, d=2, c=3". Then it would proceed as before, and obtain "4, 5, 2, 4, 5, 0, 3" with a vocabulary of "a=0, b=1, d=2, c=3, aa=4, ab=5". So far this is essentially the same as before. However, if we only had specified a vocabulary size of 5, then the process would stop at vocabulary "a=0, b=1, d=2, c=3, aa=4", so that the example would be encoded as "4, 0, 1, 2, 4, 0, 1, 0, 3". Conversely, if we had specified a vocabulary size of 8, then it would be encoded as "7, 6, 0, 3", with a vocabulary of "a=0, b=1, d=2, c=3, aa=4, ab=5, aaab=6, aaabd=7". This is not maximally compressed, because modified BPE does not aim for maximum compression. Instead, it aims for an encoding that is efficient and practical for language model training. ==See also== Re-Pair Sequitur algorithm ==References==

Text derived from https://en.wikipedia.org/w/index.php?oldid=1346605957 under CC BY-SA 4.0; modifications have been made.