CE - Wikilangs Models

Comprehensive Research Report & Full Ablation Study

This repository contains NLP models trained and evaluated by Wikilangs, specifically on CE Wikipedia data. We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and word embeddings.

📋 Repository Contents

Models & Assets

• Tokenizers (8k, 16k, 32k, 64k)
• N-gram models (2, 3, 4, 5-gram)
• Markov chains (context of 1, 2, 3, 4 and 5)
• Subword N-gram and Markov chains
• Embeddings in various sizes and dimensions (aligned and unaligned)
• Language Vocabulary
• Language Statistics

Analysis and Evaluation

• 1. Tokenizer Evaluation
• 2. N-gram Model Evaluation
• 3. Markov Chain Evaluation
• 4. Vocabulary Analysis
• 5. Word Embeddings Evaluation
• 6. Morphological Analysis (Experimental)
• 7. Summary & Recommendations
• Metrics Glossary
• Visualizations Index

---

1. Tokenizer Evaluation

Results

Vocab Size	Compression	Avg Token Len	UNK Rate	Total Tokens
8k	2.792x	2.80	0.9604%	543,837
16k	3.140x	3.15	1.0803%	483,478
32k	3.480x	3.49	1.1970%	436,328
64k	3.783x 🏆	3.79	1.3016%	401,281

Tokenization Examples

Below are sample sentences tokenized with each vocabulary size:

Sample 1: Жаныспай (Акмолан область) Жаныспай (Костанайн область)

Vocab	Tokens	Count
8k	▁жан ыс п ай ▁( ак молан ▁область ) ▁жан ... (+8 more)	18
16k	▁жан ыс пай ▁( акмолан ▁область ) ▁жан ыс пай ... (+5 more)	15
32k	▁жан ыс пай ▁( акмолан ▁область ) ▁жан ыс пай ... (+4 more)	14
64k	▁жан ыс пай ▁( акмолан ▁область ) ▁жан ыс пай ... (+4 more)	14

Sample 2: Антиго (Висконсин) Антиго (Маса-Карара) Антиго (гӀала, Висконсин)

Vocab	Tokens	Count
8k	▁анти го ▁( ви сконсин ) ▁анти го ▁( ма ... (+12 more)	22
16k	▁анти го ▁( висконсин ) ▁анти го ▁( ма са ... (+11 more)	21
32k	▁анти го ▁( висконсин ) ▁анти го ▁( маса - ... (+9 more)	19
64k	▁анти го ▁( висконсин ) ▁анти го ▁( маса - ... (+9 more)	19

Sample 3: Барда (Иркутскан область) Барда (Пермийн мохк) Барда (гӀала)

Vocab	Tokens	Count
8k	▁бар да ▁( иркутскан ▁область ) ▁бар да ▁( пермийн ... (+7 more)	17
16k	▁бар да ▁( иркутскан ▁область ) ▁бар да ▁( пермийн ... (+7 more)	17
32k	▁барда ▁( иркутскан ▁область ) ▁барда ▁( пермийн ▁мохк ) ... (+4 more)	14
64k	▁барда ▁( иркутскан ▁область ) ▁барда ▁( пермийн ▁мохк ) ... (+4 more)	14

Key Findings

• Best Compression: 64k achieves 3.783x compression
• Lowest UNK Rate: 8k with 0.9604% unknown tokens
• Trade-off: Larger vocabularies improve compression but increase model size
• Recommendation: 32k vocabulary provides optimal balance for production use

---

2. N-gram Model Evaluation

Results

N-gram	Variant	Perplexity	Entropy	Unique N-grams	Top-100 Coverage	Top-1000 Coverage
2-gram	Word	2,545	11.31	100,140	25.5%	70.0%
2-gram	Subword	423 🏆	8.72	6,176	55.1%	98.2%
3-gram	Word	3,286	11.68	157,541	21.2%	65.9%
3-gram	Subword	2,337	11.19	58,954	23.8%	69.8%
4-gram	Word	4,089	12.00	330,019	18.2%	63.2%
4-gram	Subword	5,832	12.51	337,533	16.4%	50.9%

Top 5 N-grams by Size

2-grams (Word):

Rank	N-gram	Count
1	нах беха	927,008
2	беха меттигаш	876,464
3	билгалдахарш хьажоргаш	387,483
4	климат кхузахь	294,017
5	сахьтан аса	272,866

3-grams (Word):

Rank	N-gram	Count
1	нах беха меттигаш	876,426
2	кӏоштан нах беха	256,950
3	климат кхузахь климат	254,686
4	бахархой билгалдахарш хьажоргаш	156,558
5	сахьтан аса йу	135,690

4-grams (Word):

Rank	N-gram	Count
1	кӏоштан нах беха меттигаш	256,946
2	лелаш ду сахьтан аса	134,397
3	нийса лелаш ду сахьтан	134,397
4	сахьтан аса йу utc	133,768
5	ду сахьтан аса йу	133,768

2-grams (Subword):

Rank	N-gram	Count
1	а _	8,696,976
2	. _	8,337,924
3	н _	7,066,559
4	а н	6,445,422
5	р а	5,305,199

3-grams (Subword):

Rank	N-gram	Count
1	а н _	4,127,441
2	_ — _	2,719,160
3	а ш _	1,910,774
4	и н _	1,668,837
5	а р а	1,610,648

4-grams (Subword):

Rank	N-gram	Count
1	т а н _	1,416,987
2	а х а р	1,374,119
3	. _ — _	1,045,081
4	а _ м е	1,006,220
5	_ м е т	999,858

Key Findings

• Best Perplexity: 2-gram (subword) with 423
• Entropy Trend: Decreases with larger n-grams (more predictable)
• Coverage: Top-1000 patterns cover ~51% of corpus
• Recommendation: 4-gram or 5-gram for best predictive performance

---

3. Markov Chain Evaluation

Results

Context	Variant	Avg Entropy	Perplexity	Branching Factor	Unique Contexts	Predictability
1	Word	0.6226	1.540	3.90	520,111	37.7%
1	Subword	0.9426	1.922	9.07	1,553	5.7%
2	Word	0.1849	1.137	1.44	2,019,671	81.5%
2	Subword	0.9737	1.964	7.37	14,069	2.6%
3	Word	0.0632	1.045	1.13	2,889,994	93.7%
3	Subword	0.8560	1.810	4.97	103,627	14.4%
4	Word	0.0320 🏆	1.022	1.08	3,246,178	96.8%
4	Subword	0.7168	1.643	3.27	515,118	28.3%

Generated Text Samples (Word-based)

Below are text samples generated from each word-based Markov chain model:

Context Size 1:

1. а ду йалташ хастоьмаш малхбален кӏошташкара пачхьалкхан европин дехьайолуш алсама гӏийлачу мехца бек...
2. нах беха меттигаш нах беха меттигаш провинцин нах беха меттигаш кӏоштан нах беха меттигаш воеводалли...
3. беха меттигаш штатан йукъахь квинс университет им м в пономарёва м прохоров т 82 т и

Context Size 2:

1. нах беха меттигаш нах беха меттигаш нисйина нах беха меттигаш кӏоштан нах беха меттигаш кӏоштан нах ...
2. беха меттигаш кӏоштан нах беха меттигаш кӏоштан нах беха меттигаш воеводаллин нах беха меттигаш нах ...
3. билгалдахарш хьажоргаш черкассин областан индексаш кӏоштан нах беха меттигаш микрокӏошташ нах беха м...

Context Size 3:

1. нах беха меттигаш микрокӏошташ нах беха меттигаш нисйина нах беха меттигаш нах беха меттигаш микрокӏ...
2. кӏоштан нах беха меттигаш нисйина нах беха меттигаш нах беха меттигаш кӏоштан нах беха меттигаш нах ...
3. климат кхузахь климат барамехь континенталан йу аьхка йовха хуьлу ткъа ӏа барамехь шийла хуьлу шаран...

Context Size 4:

1. нийса лелаш ду сахьтан аса йу utc 3 билгалдахарш хьажоргаш неклиновскан кӏоштан индексаш кӏоштан нах...
2. лелаш ду сахьтан аса йу utc 3 билгалдахарш хьажоргаш селижарован кӏоштан индексаш кӏоштан нах беха м...
3. ду сахьтан аса йу utc 3 билгалдахарш хьажоргаш максатихан кӏоштан индексаш кӏоштан нах беха меттигаш...

Generated Text Samples (Subword-based)

Below are text samples generated from each subword-based Markov chain model:

Context Size 1:

1. _/_циллан_olevia
2. а_перерхашес._ба
3. нилию._7959-со_к

Context Size 2:

1. а_койн_сахар_тӏуь
2. ._ре_нашкая_:_спу
3. н_схойн_стр_штаме

Context Size 3:

1. ан_аркатерия_исти_
2. _—_итан_новгорокӏо
3. аш_беха_местник_гу

Context Size 4:

1. тан_кӏоштан_воеводс
2. ахарш_хьажоргаш_нах
3. ._—_b.,_heidelberg,

Key Findings

• Best Predictability: Context-4 (word) with 96.8% predictability
• Branching Factor: Decreases with context size (more deterministic)
• Memory Trade-off: Larger contexts require more storage (515,118 contexts)
• Recommendation: Context-3 or Context-4 for text generation

---

4. Vocabulary Analysis

Statistics

Metric	Value
Vocabulary Size	230,774
Total Tokens	54,539,322
Mean Frequency	236.33
Median Frequency	3
Frequency Std Dev	7087.98

Most Common Words

Rank	Word	Frequency
1	а	1,429,788
2	нах	929,389
3	беха	927,412
4	меттигаш	892,206
5	в	665,820
6	климат	663,481
7	м	649,926
8	йу	631,461
9	билгалдахарш	595,304
10	с	497,975

Least Common Words (from vocabulary)

Rank	Word	Frequency
1	горушкинскан	2
2	тулинскан	2
3	долгопольскан	2
4	погостищенскан	2
5	кохановскан	2
6	морховскан	2
7	нежадовскан	2
8	липиницкан	2
9	зачепичи	2
10	меетиг	2

Zipf's Law Analysis

Metric	Value
Zipf Coefficient	1.8318
R² (Goodness of Fit)	0.964473
Adherence Quality	excellent

Coverage Analysis

Top N Words	Coverage
Top 100	44.4%
Top 1,000	86.7%
Top 5,000	96.7%
Top 10,000	97.7%

Key Findings

• Zipf Compliance: R²=0.9645 indicates excellent adherence to Zipf's law
• High Frequency Dominance: Top 100 words cover 44.4% of corpus
• Long Tail: 220,774 words needed for remaining 2.3% coverage

---

5. Word Embeddings Evaluation

5.1 Cross-Lingual Alignment

Note: Multilingual alignment visualization not available for this language.

5.2 Model Comparison

Model	Dimension	Isotropy	Semantic Density	Alignment R@1	Alignment R@10
mono_32d	32	0.8761 🏆	0.3710	N/A	N/A
mono_64d	64	0.8520	0.3045	N/A	N/A
mono_128d	128	0.7849	0.2825	N/A	N/A

Key Findings

• Best Isotropy: mono_32d with 0.8761 (more uniform distribution)
• Semantic Density: Average pairwise similarity of 0.3193. Lower values indicate better semantic separation.
• Alignment Quality: No aligned models evaluated in this run.
• Recommendation: 128d aligned for best cross-lingual performance

---

6. Morphological Analysis (Experimental)

⚠️ Warning: This language shows low morphological productivity. The statistical signals used for this analysis may be noisy or less reliable than for morphologically rich languages.

This section presents an automated morphological analysis derived from the statistical divergence between word-level and subword-level models. By analyzing where subword predictability spikes and where word-level coverage fails, we can infer linguistic structures without supervised data.

6.1 Productivity & Complexity

Metric	Value	Interpretation	Recommendation
Productivity Index	0.000	Low morphological productivity	⚠️ Likely unreliable
Idiomaticity Gap	-1.000	Low formulaic content	-

6.2 Affix Inventory (Productive Units)

These are the most productive prefixes and suffixes identified by sampling the vocabulary for global substitutability patterns. A unit is considered an affix if stripping it leaves a valid stem that appears in other contexts.

Productive Prefixes

Prefix	Examples
-ка	картографии, карайора, карпат
-ко	количество, кочаны, кошехаблан
-ма	майкен, маршаллвилл, машано

Productive Suffixes

Suffix	Examples
-а	ривица, валенсуэла, карайора
-о	монтеморо, количество, мятнево
-н	расистийн, майкен, тефран
-ан	тефран, дмитрован, кертан
-во	количество, мятнево, крайково
-ки	исаковски, юридически, перлавки
-ово	крайково, перегудово, дубново
-ка	узника, кукушка, тлаика

6.3 Bound Stems (Lexical Roots)

Bound stems are high-frequency subword units that are semantically cohesive but rarely appear as standalone words. These often correspond to the 'core' of a word that requires inflection or derivation to be valid.

Stem	Cohesion	Substitutability	Examples
архо	2.04x	122 contexts	архон, тархо, лархо
галд	2.73x	16 contexts	галдо, галда, угалде
ргаш	2.16x	34 contexts	ургаш, бергаш, цергаш
лгал	2.58x	17 contexts	билгал, билгало, билгала
етти	1.89x	42 contexts	бетти, нетти, меттин
харх	1.88x	41 contexts	ахархо, вахарх, мухарх
халл	1.51x	92 contexts	халла, халле, халль
ийла	1.86x	35 contexts	кийла, шийла, мийла
игаш	2.25x	18 contexts	бигаш, цигаш, книгаш
рхой	2.21x	19 contexts	лархой, сурхой, сурхойн
ласт	1.59x	60 contexts	пласт, ласта, селаст
ттиг	1.99x	25 contexts	меттиг, гаттиг, ме́ттиг

6.4 Affix Compatibility (Co-occurrence)

This table shows which prefixes and suffixes most frequently co-occur on the same stems, revealing the 'stacking' rules of the language's morphology.

Prefix	Suffix	Frequency	Examples
-ко	-а	51 words	королиха, кокориха
-ка	-а	43 words	карпеевка, камила
-ка	-о	38 words	картелево, катюшино
-ма	-а	35 words	машакепара, малакода
-ко	-о	33 words	косогорово, косяково
-ка	-н	31 words	калустовгӏеран, камблен
-ма	-н	24 words	малоярославецан, марьинкан
-ко	-н	23 words	коритен, койдин
-ма	-о	22 words	маторо, манкузо
-ко	-во	18 words	косогорово, косяково

6.5 Recursive Morpheme Segmentation

Using Recursive Hierarchical Substitutability, we decompose complex words into their constituent morphemes. This approach handles nested affixes (e.g., prefix-prefix-root-suffix).

Word	Suggested Split	Confidence	Stem
полканово	полк-ан-ово	6.0	полк
андрюшкино	андрюш-ки-но	6.0	андрюш
зимовники	зимовни-ки	4.5	зимовни
гринвичан	гринвич-ан	4.5	гринвич
гуннбьёрнан	гуннбьёрн-ан	4.5	гуннбьёрн
хромосоман	хромосом-ан	4.5	хромосом
боьлкъазаран	боьлкъазар-ан	4.5	боьлкъазар
хӏуманашна	хӏуманаш-на	4.5	хӏуманаш
ынтымакан	ынтымак-ан	4.5	ынтымак
бартолина	бартоли-на	4.5	бартоли
судженскан	судженск-ан	4.5	судженск
бузиновка	бузинов-ка	4.5	бузинов
тракторашна	трактораш-на	4.5	трактораш
пайхӏамаран	пайхӏамар-ан	4.5	пайхӏамар
нуьрнберган	нуьрнберг-ан	4.5	нуьрнберг

6.6 Linguistic Interpretation

Automated Insight: The language CE appears to be more isolating or has a highly fixed vocabulary. Word-level models perform nearly as well as subword models, indicating fewer productive morphological processes.

---

7. Summary & Recommendations

Production Recommendations

Component	Recommended	Rationale
Tokenizer	64k BPE	Best compression (3.78x)
N-gram	2-gram	Lowest perplexity (423)
Markov	Context-4	Highest predictability (96.8%)
Embeddings	100d	Balanced semantic capture and isotropy

---

Appendix: Metrics Glossary & Interpretation Guide

This section provides definitions, intuitions, and guidance for interpreting the metrics used throughout this report.

Tokenizer Metrics

Compression Ratio

Definition: The ratio of characters to tokens (chars/token). Measures how efficiently the tokenizer represents text.

Intuition: Higher compression means fewer tokens needed to represent the same text, reducing sequence lengths for downstream models. A 3x compression means ~3 characters per token on average.

What to seek: Higher is generally better for efficiency, but extremely high compression may indicate overly aggressive merging that loses morphological information.

Average Token Length (Fertility)

Definition: Mean number of characters per token produced by the tokenizer.

Intuition: Reflects the granularity of tokenization. Longer tokens capture more context but may struggle with rare words; shorter tokens are more flexible but increase sequence length.

What to seek: Balance between 2-5 characters for most languages. Arabic/morphologically-rich languages may benefit from slightly longer tokens.

Unknown Token Rate (OOV Rate)

Definition: Percentage of tokens that map to the unknown/UNK token, indicating words the tokenizer cannot represent.

Intuition: Lower OOV means better vocabulary coverage. High OOV indicates the tokenizer encounters many unseen character sequences.

What to seek: Below 1% is excellent; below 5% is acceptable. BPE tokenizers typically achieve very low OOV due to subword fallback.

N-gram Model Metrics

Perplexity

Definition: Measures how "surprised" the model is by test data. Mathematically: 2^(cross-entropy). Lower values indicate better prediction.

Intuition: If perplexity is 100, the model is as uncertain as if choosing uniformly among 100 options at each step. A perplexity of 10 means effectively choosing among 10 equally likely options.

What to seek: Lower is better. Perplexity decreases with larger n-grams (more context). Values vary widely by language and corpus size.

Entropy

Definition: Average information content (in bits) needed to encode the next token given the context. Related to perplexity: perplexity = 2^entropy.

Intuition: High entropy means high uncertainty/randomness; low entropy means predictable patterns. Natural language typically has entropy between 1-4 bits per character.

What to seek: Lower entropy indicates more predictable text patterns. Entropy should decrease as n-gram size increases.

Coverage (Top-K)

Definition: Percentage of corpus occurrences explained by the top K most frequent n-grams.

Intuition: High coverage with few patterns indicates repetitive/formulaic text; low coverage suggests diverse vocabulary usage.

What to seek: Depends on use case. For language modeling, moderate coverage (40-60% with top-1000) is typical for natural text.

Markov Chain Metrics

Average Entropy

Definition: Mean entropy across all contexts, measuring average uncertainty in next-word prediction.

Intuition: Lower entropy means the model is more confident about what comes next. Context-1 has high entropy (many possible next words); Context-4 has low entropy (few likely continuations).

What to seek: Decreasing entropy with larger context sizes. Very low entropy (<0.1) indicates highly deterministic transitions.

Branching Factor

Definition: Average number of unique next tokens observed for each context.

Intuition: High branching = many possible continuations (flexible but uncertain); low branching = few options (predictable but potentially repetitive).

What to seek: Branching factor should decrease with context size. Values near 1.0 indicate nearly deterministic chains.

Predictability

Definition: Derived metric: (1 - normalized_entropy) × 100%. Indicates how deterministic the model's predictions are.

Intuition: 100% predictability means the next word is always certain; 0% means completely random. Real text falls between these extremes.

What to seek: Higher predictability for text generation quality, but too high (>98%) may produce repetitive output.

Vocabulary & Zipf's Law Metrics

Zipf's Coefficient

Definition: The slope of the log-log plot of word frequency vs. rank. Zipf's law predicts this should be approximately -1.

Intuition: A coefficient near -1 indicates the corpus follows natural language patterns where a few words are very common and most words are rare.

What to seek: Values between -0.8 and -1.2 indicate healthy natural language distribution. Deviations may suggest domain-specific or artificial text.

R² (Coefficient of Determination)

Definition: Measures how well the linear fit explains the frequency-rank relationship. Ranges from 0 to 1.

Intuition: R² near 1.0 means the data closely follows Zipf's law; lower values indicate deviation from expected word frequency patterns.

What to seek: R² > 0.95 is excellent; > 0.99 indicates near-perfect Zipf adherence typical of large natural corpora.

Vocabulary Coverage

Definition: Cumulative percentage of corpus tokens accounted for by the top N words.

Intuition: Shows how concentrated word usage is. If top-100 words cover 50% of text, the corpus relies heavily on common words.

What to seek: Top-100 covering 30-50% is typical. Higher coverage indicates more repetitive text; lower suggests richer vocabulary.

Word Embedding Metrics

Isotropy

Definition: Measures how uniformly distributed vectors are in the embedding space. Computed as the ratio of minimum to maximum singular values.

Intuition: High isotropy (near 1.0) means vectors spread evenly in all directions; low isotropy means vectors cluster in certain directions, reducing expressiveness.

What to seek: Higher isotropy generally indicates better-quality embeddings. Values > 0.1 are reasonable; > 0.3 is good. Lower-dimensional embeddings tend to have higher isotropy.

Average Norm

Definition: Mean magnitude (L2 norm) of word vectors in the embedding space.

Intuition: Indicates the typical "length" of vectors. Consistent norms suggest stable training; high variance may indicate some words are undertrained.

What to seek: Relatively consistent norms across models. The absolute value matters less than consistency (low std deviation).

Cosine Similarity

Definition: Measures angular similarity between vectors, ranging from -1 (opposite) to 1 (identical direction).

Intuition: Words with similar meanings should have high cosine similarity. This is the standard metric for semantic relatedness in embeddings.

What to seek: Semantically related words should score > 0.5; unrelated words should be near 0. Synonyms often score > 0.7.

t-SNE Visualization

Definition: t-Distributed Stochastic Neighbor Embedding - a dimensionality reduction technique that preserves local structure for visualization.

Intuition: Clusters in t-SNE plots indicate groups of semantically related words. Spread indicates vocabulary diversity; tight clusters suggest semantic coherence.

What to seek: Meaningful clusters (e.g., numbers together, verbs together). Avoid over-interpreting distances - t-SNE preserves local, not global, structure.

General Interpretation Guidelines

1. Compare within model families: Metrics are most meaningful when comparing models of the same type (e.g., 8k vs 64k tokenizer).
2. Consider trade-offs: Better performance on one metric often comes at the cost of another (e.g., compression vs. OOV rate).
3. Context matters: Optimal values depend on downstream tasks. Text generation may prioritize different metrics than classification.
4. Corpus influence: All metrics are influenced by corpus characteristics. Wikipedia text differs from social media or literature.
5. Language-specific patterns: Morphologically rich languages (like Arabic) may show different optimal ranges than analytic languages.

Visualizations Index

Visualization	Description
Tokenizer Compression	Compression ratios by vocabulary size
Tokenizer Fertility	Average token length by vocabulary
Tokenizer OOV	Unknown token rates
Tokenizer Total Tokens	Total tokens by vocabulary
N-gram Perplexity	Perplexity by n-gram size
N-gram Entropy	Entropy by n-gram size
N-gram Coverage	Top pattern coverage
N-gram Unique	Unique n-gram counts
Markov Entropy	Entropy by context size
Markov Branching	Branching factor by context
Markov Contexts	Unique context counts
Zipf's Law	Frequency-rank distribution with fit
Vocab Frequency	Word frequency distribution
Top 20 Words	Most frequent words
Vocab Coverage	Cumulative coverage curve
Embedding Isotropy	Vector space uniformity
Embedding Norms	Vector magnitude distribution
Embedding Similarity	Word similarity heatmap
Nearest Neighbors	Similar words for key terms
t-SNE Words	2D word embedding visualization
t-SNE Sentences	2D sentence embedding visualization
Position Encoding	Encoding method comparison
Model Sizes	Storage requirements
Performance Dashboard	Comprehensive performance overview

---

About This Project

Data Source

Models trained on wikipedia-monthly - a monthly snapshot of Wikipedia articles across 300+ languages.

Project

A project by Wikilangs - Open-source NLP models for every Wikipedia language.

Maintainer

Omar Kamali - Omneity Labs

Citation

If you use these models in your research, please cite:

@misc{wikilangs2025,
  author = {Kamali, Omar},
  title = {Wikilangs: Open NLP Models for Wikipedia Languages},
  year = {2025},
  doi = {10.5281/zenodo.18073153},
  publisher = {Zenodo},
  url = {https://huggingface.co/wikilangs}
  institution = {Omneity Labs}
}

License

MIT License - Free for academic and commercial use.

Links

• 🌐 Website: wikilangs.org
• 🤗 Models: huggingface.co/wikilangs
• 📊 Data: wikipedia-monthly
• 👤 Author: Omar Kamali
• 🤝 Sponsor: Featherless AI

---

Generated by Wikilangs Models Pipeline

Report Date: 2026-01-03 10:17:57