LM filter with better support for cyrillic

lm.en

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:caae8bb4c4beede9b29665f4ca9f76e1248c4750b1e3e3d3c8516d30ad22c471
-size 171883363
+oid sha256:8a5cb9a347085a461ecd67bb4c6ac6c4dc5a8423a5b6854dac29d72b7faed6a1
+size 90786329

lm.hbs

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:931d0fb2ea4851a3a27a9d97949066aa28d60739925cee573ddb52bd81045fe7
-size 207065876
+oid sha256:4e045c50716156ed186865db8874df3a2144e03f55b48fa89343efc47df68678
+size 343857442

metadata.yaml

@@ -10,10 +10,10 @@ porn_removal_side: sl
 source_lm: lm.en
 target_lm: lm.hbs
 lm_type: CHARACTER
-clean_mean_perp: -1.095718928970861
-clean_stddev_perp: 0.17303607985831154
-noisy_mean_perp: -4.0844201427926
-noisy_stddev_perp: 0.17576658838565767
+clean_mean_perp: -1.116365617208272
+clean_stddev_perp: 0.1687608181840407
+noisy_mean_perp: -4.013399029709761
+noisy_stddev_perp: 0.15687448432446158
 classifier_type: xlmr
 classifier_settings:
     activation: relu