# m.model is a BPE (not unigram lm!) model from the sentence piece library 
# Note: for unigram lm model see <BlingFire>/ldbsrc/xlnet and <BlingFire>/ldbsrc/xlnet _nonorm examples

# export the model:
spm_export_vocab --model m.model --output spiece.model.exportvocab.txt --output_format txt 

# produce pos.dict.utf8 file and tagset.txt:
cat spiece.model.exportvocab.txt | awk "BEGIN {FS="\t"} NF == 2 { if (NR > 1) { print $1 "\tWORD_ID_" NR-1 "\t" ($2 == 0 ? "-0.00001" : $2); }  print "WORD_ID_" NR " " NR > "tagset.txt"; }" > pos.dict.utf8

# zip it:
zip pos.dict.utf8.zip pos.dict.utf8

# build as usual
make -f Makefile.gnu lang=bpe_example all

# test parity between m.model and bpe_example.bin :
> cat input.utf8 | python ../scripts/test_bling_with_offsets.py -m bpe_example.bin -p m.model -u 3 -k 1 | awk "/ERROR/" | wc -l
3881

> wc -l input.utf8
2052515 input.utf8

99.8% parity



Now let"s measure performance and compare it to the sentencepiece library:

> time -p cat input.utf8 | python ../scripts/test_bling.py -m bpe_example.bin -s 1 
real 125.22
user 124.99
sys 1.74

> time -p cat input.utf8 | python ../scripts/test_sp.py -m m.model -s 1 
real 262.10
user 261.50
sys 1.68

As you can see Bling Fire BPE (optimized) runs ~2x faster than the sentence piece library.