ban
Balinese
Performance Dashboard
Quick Start
Tokenizer
from wikilangs import tokenizer
tok = tokenizer(date='latest', lang='ban', vocab_size=32000)
tokens = tok.tokenize("Your text here")
print(tokens)N-gram Model
from wikilangs import ngram
ng = ngram(date='latest', lang='ban', gram_size=3)
score = ng.score("Your text here")
predictions = ng.predict_next("Start of", top_k=5)Markov Chain
from wikilangs import markov
mc = markov(date='latest', lang='ban', depth=3)
text = mc.generate(length=50)
print(text)Vocabulary
from wikilangs import vocabulary
vocab = vocabulary(date='latest', lang='ban')
info = vocab.lookup("word")
print(info) # frequency, IDF, rankEmbeddings
from wikilangs import embeddings
emb = embeddings(date='latest', lang='ban', dimension=64)
vec = emb.embed_word("word")
sent_vec = emb.embed_sentence("A sentence", method='rope')Available Models
| Model Type | Variants | Description |
|---|---|---|
| Tokenizers | 8k, 16k, 32k, 64k | BPE tokenizers with different vocabulary sizes |
| N-gram (Word) | 2, 3, 4, 5-gram | Word-level language models |
| N-gram (Subword) | 2, 3, 4, 5-gram | Subword-level language models |
| Markov (Word) | Depth 1-5 | Word-level text generation |
| Markov (Subword) | Depth 1-5 | Subword-level text generation |
| Vocabulary | — | Word dictionary with frequency and IDF |
| Embeddings | 32d, 64d, 128d | Position-aware word embeddings |
Model Evaluation
Tokenizer Performance
Compression ratios and token statistics across vocabulary sizes. Higher compression means fewer tokens for the same text.
N-gram Model Evaluation
Perplexity and entropy metrics across n-gram sizes. Lower perplexity indicates better predictive performance.
Markov Chain Evaluation
Entropy and branching factor by context depth. Lower entropy means more predictable text generation.
Vocabulary Analysis
Word frequency distribution and Zipf's law analysis.
Embeddings Evaluation
Isotropy and vector space quality metrics. Higher isotropy indicates more uniformly distributed embeddings.
Full Research Report
Access the complete ablation study with all metrics, visualizations, and generated text samples on HuggingFace.
View Full Report on HuggingFace