BoKenLm - Tibetan KenLM Language Model

A KenLM n-gram language model trained on Tibetan text, tokenized with BoSentencePiece.

Model Details

Parameter	Value
Model Type	Modified Kneser-Ney 5-gram
Tokenizer	openpecha/BoSentencePiece (Unigram, 20k vocab)
Training Corpus	bo_corpus.txt
Pruning	0 0 1
Tokens	38,497,401
Vocabulary Size	20,003

N-gram Statistics

Order	Count	D1	D2	D3+
1	20,003	0.4380	0.4927	1.5624
2	6,649,657	0.6717	1.1476	1.5422
3	4,299,504	0.8465	1.2657	1.4805
4	3,477,865	0.9176	1.3860	1.5187
5	2,589,246	0.8776	1.4493	1.5850

Memory Estimates

Type	MB	Details
probing	375	assuming -p 1.5
probing	457	assuming -r models -p 1.5
trie	187	without quantization
trie	99	assuming -q 8 -b 8 quantization
trie	159	assuming -a 22 array pointer compression
trie	71	assuming -a 22 -q 8 -b 8 array pointer compression and quantization

Training Resources

Metric	Value
Peak Virtual Memory	12,333 MB
Peak RSS	2,976 MB
Wall Time	36.2s
User Time	41.3s
System Time	17.1s

Usage

import kenlm

model = kenlm.Model("lm.arpa")

# Score a tokenized sentence
score = model.score("▁བོད་སྐད་ ▁ཀྱི་ ▁ཚིག་གྲུབ་ ▁འདི་ ▁ཡིན།")
print(score)

Files

• lm.arpa — ARPA format language model
• README.md — This model card

License

Apache 2.0