wikilangs/ady
0
1---2language: ady3language_name: Adyghe4language_family: caucasian_northwest5tags:6 - wikilangs7 - nlp8 - tokenizer9 - embeddings10 - n-gram11 - markov12 - wikipedia13 - feature-extraction14 - sentence-similarity15 - tokenization16 - n-grams17 - markov-chain18 - text-mining19 - fasttext20 - babelvec21 - vocabulous22 - vocabulary23 - monolingual24 - family-caucasian_northwest25license: mit26library_name: wikilangs27pipeline_tag: text-generation28datasets:29 - omarkamali/wikipedia-monthly30dataset_info:31 name: wikipedia-monthly32 description: Monthly snapshots of Wikipedia articles across 300+ languages33metrics:34 - name: best_compression_ratio35 type: compression36 value: 4.19737 - name: best_isotropy38 type: isotropy39 value: 0.488040 - name: vocabulary_size41 type: vocab42 value: 043generated: 2026-01-0344---45 46# Adyghe - Wikilangs Models47## Comprehensive Research Report & Full Ablation Study48 49This repository contains NLP models trained and evaluated by Wikilangs, specifically on **Adyghe** Wikipedia data.50We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and word embeddings.51 52## 📋 Repository Contents53 54### Models & Assets55 56- Tokenizers (8k, 16k, 32k, 64k)57- N-gram models (2, 3, 4, 5-gram)58- Markov chains (context of 1, 2, 3, 4 and 5)59- Subword N-gram and Markov chains60- Embeddings in various sizes and dimensions (aligned and unaligned)61- Language Vocabulary62- Language Statistics63 6465 66### Analysis and Evaluation67 68- [1. Tokenizer Evaluation](#1-tokenizer-evaluation)69- [2. N-gram Model Evaluation](#2-n-gram-model-evaluation)70- [3. Markov Chain Evaluation](#3-markov-chain-evaluation)71- [4. Vocabulary Analysis](#4-vocabulary-analysis)72- [5. Word Embeddings Evaluation](#5-word-embeddings-evaluation)73- [6. Morphological Analysis (Experimental)](#6--morphological-analysis-experimental)74- [7. Summary & Recommendations](#7-summary--recommendations)75- [Metrics Glossary](#appendix-metrics-glossary--interpretation-guide)76- [Visualizations Index](#visualizations-index)77 78---79## 1. Tokenizer Evaluation80 8182 8384 8586 8788 89### Results90 91| Vocab Size | Compression | Avg Token Len | UNK Rate | Total Tokens |92|------------|-------------|---------------|----------|--------------|93| **8k** | 3.406x | 3.41 | 0.1685% | 137,125 |94| **16k** | 3.759x | 3.76 | 0.1859% | 124,248 |95| **32k** | 4.197x 🏆 | 4.20 | 0.2076% | 111,273 |96 97### Tokenization Examples98 99Below are sample sentences tokenized with each vocabulary size:100 101**Sample 1:** `Ермэлхэр — Кавказым ыкӏи дунаем тет лъэпкъ жъыдэдэмэ ащыщых. Армение`102 103| Vocab | Tokens | Count |104|-------|--------|-------|105| 8k | `▁ермэлхэр ▁— ▁кавказым ▁ыкӏи ▁дунаем ▁тет ▁лъэпкъ ▁жъыдэдэмэ ▁ащыщых . ... (+1 more)` | 11 |106| 16k | `▁ермэлхэр ▁— ▁кавказым ▁ыкӏи ▁дунаем ▁тет ▁лъэпкъ ▁жъыдэдэмэ ▁ащыщых . ... (+1 more)` | 11 |107| 32k | `▁ермэлхэр ▁— ▁кавказым ▁ыкӏи ▁дунаем ▁тет ▁лъэпкъ ▁жъыдэдэмэ ▁ащыщых . ... (+1 more)` | 11 |108 109**Sample 2:** `ТӀэшъу Светлан (УрысыбзэкӀэ: Светлана Тешева) Адыгэ журналист Адыгеим щыщ.`110 111| Vocab | Tokens | Count |112|-------|--------|-------|113| 8k | `▁тӏэ шъу ▁светлан ▁( урысыбзэкӏэ : ▁светлан а ▁те ше ... (+7 more)` | 17 |114| 16k | `▁тӏэ шъу ▁светлан ▁( урысыбзэкӏэ : ▁светлана ▁тешева ) ▁адыгэ ... (+4 more)` | 14 |115| 32k | `▁тӏэшъу ▁светлан ▁( урысыбзэкӏэ : ▁светлана ▁тешева ) ▁адыгэ ▁журналист ... (+3 more)` | 13 |116 117**Sample 3:** `Ашрай - быслъымэнмэ къурмэным ыуж мэфэ гъэнэфагъэм щагъэжъорэ стырыпс. category`118 119| Vocab | Tokens | Count |120|-------|--------|-------|121| 8k | `▁аш рай ▁- ▁быслъымэн мэ ▁къур мэным ▁ыуж ▁мэфэ ▁гъэнэф ... (+9 more)` | 19 |122| 16k | `▁аш рай ▁- ▁быслъымэн мэ ▁къурмэным ▁ыуж ▁мэфэ ▁гъэнэфагъэм ▁щагъэ ... (+4 more)` | 14 |123| 32k | `▁ашрай ▁- ▁быслъымэнмэ ▁къурмэным ▁ыуж ▁мэфэ ▁гъэнэфагъэм ▁щагъэжъорэ ▁стырыпс . ... (+1 more)` | 11 |124 125 126### Key Findings127 128- **Best Compression:** 32k achieves 4.197x compression129- **Lowest UNK Rate:** 8k with 0.1685% unknown tokens130- **Trade-off:** Larger vocabularies improve compression but increase model size131- **Recommendation:** 32k vocabulary provides optimal balance for production use132 133---134## 2. N-gram Model Evaluation135 136137 138139 140141 142### Results143 144| N-gram | Variant | Perplexity | Entropy | Unique N-grams | Top-100 Coverage | Top-1000 Coverage |145|--------|---------|------------|---------|----------------|------------------|-------------------|146| **2-gram** | Word | 453 | 8.82 | 625 | 42.6% | 100.0% |147| **2-gram** | Subword | 407 🏆 | 8.67 | 2,126 | 56.6% | 97.3% |148| **3-gram** | Word | 759 | 9.57 | 977 | 31.6% | 100.0% |149| **3-gram** | Subword | 2,854 | 11.48 | 11,856 | 24.3% | 64.6% |150| **4-gram** | Word | 2,909 | 11.51 | 3,378 | 13.2% | 45.0% |151| **4-gram** | Subword | 10,911 | 13.41 | 36,062 | 12.4% | 39.2% |152| **5-gram** | Word | 2,658 | 11.38 | 2,950 | 12.2% | 45.6% |153| **5-gram** | Subword | 21,199 | 14.37 | 52,393 | 8.2% | 28.3% |154 155### Top 5 N-grams by Size156 157**2-grams (Word):**158 159| Rank | N-gram | Count |160|------|--------|-------|161| 1 | `нэбгырэ млн` | 168 |162| 2 | `къехъу щэпсэу` | 104 |163| 3 | `м къехъу` | 89 |164| 4 | `дло м` | 87 |165| 5 | `адыгэ республикэм` | 80 |166 167**3-grams (Word):**168 169| Rank | N-gram | Count |170|------|--------|-------|171| 1 | `м къехъу щэпсэу` | 76 |172| 2 | `къехъу щэпсэу хэгэгум` | 70 |173| 3 | `адыгэ республикэм и` | 46 |174| 4 | `дло м хахьэ` | 44 |175| 5 | `м хахьэ хэгъэгу` | 39 |176 177**4-grams (Word):**178 179| Rank | N-gram | Count |180|------|--------|-------|181| 1 | `м къехъу щэпсэу хэгэгум` | 45 |182| 2 | `дло м хахьэ хэгъэгу` | 39 |183| 3 | `еуропэм хэт къэралыгъу къэлэ` | 23 |184| 4 | `америкэм ит къэралыгъу къэлэ` | 19 |185| 5 | `азием ит къэралыгъу къэлэ` | 18 |186 187**5-grams (Word):**188 189| Rank | N-gram | Count |190|------|--------|-------|191| 1 | `км гъогу щыӏ къуаджэм ис` | 17 |192| 2 | `гъогу щыӏ къуаджэм ис цӏыфхэр` | 17 |193| 3 | `щыӏ къуаджэм ис цӏыфхэр илъэсхэм` | 17 |194| 4 | `къуаджэм ис цӏыфхэр илъэсхэм тетэу` | 17 |195| 5 | `ис цӏыфхэр илъэсхэм тетэу къуаджэм` | 17 |196 197**2-grams (Subword):**198 199| Rank | N-gram | Count |200|------|--------|-------|201| 1 | `г ъ` | 9,326 |202| 2 | `ъ э` | 9,249 |203| 3 | `э _` | 8,792 |204| 4 | `м _` | 7,740 |205| 5 | `э р` | 6,822 |206 207**3-grams (Subword):**208 209| Rank | N-gram | Count |210|------|--------|-------|211| 1 | `г ъ э` | 4,961 |212| 2 | `_ к ъ` | 4,140 |213| 3 | `э м _` | 3,581 |214| 4 | `ы г ъ` | 3,362 |215| 5 | `э р _` | 3,020 |216 217**4-grams (Subword):**218 219| Rank | N-gram | Count |220|------|--------|-------|221| 1 | `ы г ъ э` | 1,902 |222| 2 | `х э р _` | 1,448 |223| 3 | `а г ъ э` | 1,342 |224| 4 | `х э м _` | 1,303 |225| 5 | `_ к ъ э` | 1,289 |226 227**5-grams (Subword):**228 229| Rank | N-gram | Count |230|------|--------|-------|231| 1 | `_ а д ы г` | 1,062 |232| 2 | `а д ы г э` | 978 |233| 3 | `_ и л ъ э` | 670 |234| 4 | `д ы г э _` | 651 |235| 5 | `и л ъ э с` | 627 |236 237 238### Key Findings239 240- **Best Perplexity:** 2-gram (subword) with 407241- **Entropy Trend:** Decreases with larger n-grams (more predictable)242- **Coverage:** Top-1000 patterns cover ~28% of corpus243- **Recommendation:** 4-gram or 5-gram for best predictive performance244 245---246## 3. Markov Chain Evaluation247 248249 250251 252253 254### Results255 256| Context | Variant | Avg Entropy | Perplexity | Branching Factor | Unique Contexts | Predictability |257|---------|---------|-------------|------------|------------------|-----------------|----------------|258| **1** | Word | 0.4341 | 1.351 | 2.09 | 22,655 | 56.6% |259| **1** | Subword | 1.4193 | 2.674 | 10.02 | 450 | 0.0% |260| **2** | Word | 0.0766 | 1.055 | 1.12 | 46,851 | 92.3% |261| **2** | Subword | 1.1376 | 2.200 | 5.57 | 4,503 | 0.0% |262| **3** | Word | 0.0248 | 1.017 | 1.04 | 51,794 | 97.5% |263| **3** | Subword | 0.7466 | 1.678 | 2.95 | 25,044 | 25.3% |264| **4** | Word | 0.0130 🏆 | 1.009 | 1.02 | 53,002 | 98.7% |265| **4** | Subword | 0.4264 | 1.344 | 1.85 | 73,859 | 57.4% |266 267### Generated Text Samples (Word-based)268 269Below are text samples generated from each word-based Markov chain model:270 271**Context Size 1:**272 2731. `и дгъэпсыфынущ адыгэ лъэпкъым и 29 м н ф ф ф ф х х х хъ`2742. `адыгэ хэхэсхэм ащыухъумэн ылъэкӏыгъ мыхъугъэ мышӏагъэхэр ыгу ит тарихъ лъапсэ иӏэу кӏэхьапӏэр ӏатау ...`2753. `м ахахьэ хэгъэгу тхьаматэр халед бахах географие еуропэм ыгу рихь римыхьмэ тетэу къуаджэм ис цӏыфхэр...`276 277**Context Size 2:**278 2791. `нэбгырэ млн 1 3 фэдиз ц1ыфэу дэс ау хьанэгъунэр ибгъэгъусэжьмэ млн 18 фэдиз мэхъу щыпсэухэрэм ромэ к...`2802. `къехъу щэпсэу я 67 норвегыбз дло м ахахьэ хэгъэгу эдгар ринкевичс къэрал тхьаматэр ульф кристерссон ...`2813. `м къехъу щэпсэу хэгэгум 1 240 192 км францыбзэ къэрал яйи бони хэгъэгу тхьаматэр халифа бен салман`282 283**Context Size 3:**284 2851. `м къехъу щэпсэу хэгэгум 147 570 км бенгалыбзэ дло м хахьэ хэгъэгу абдель азиз бутефлика къэрал тхьэм...`2862. `къехъу щэпсэу хэгэгум 140 800 км непали дло м хахьэ ез м хэхьанэу унашъо щыт ез м и`2873. `адыгэ республикэм и псыхъу а псыхъом пэблагъэу щыт къуажэ`288 289**Context Size 4:**290 2911. `м къехъу щэпсэу хэгэгум чӏырэу иӏэр 322 460 км бзэшъхьаӏэхэр францыбзэ къэрал лӏышъхьэр алассан уатт...`2922. `дло м хахьэ хэгъэгу султанэу кабоос бин саид аль саид хэгъэгу тхьаматэр фахд бин махьмуд географие а...`2933. `еуропэм хэт къэралыгъу къэлэ тирана нэбгырэ млн 3 м къехъу щэпсэу хэгэгум 9 984 670 км я 2 англыбзэ`294 295 296### Generated Text Samples (Subword-based)297 298Below are text samples generated from each subword-based Markov chain model:299 300**Context Size 1:**301 3021. `_шхажъырэм_ащтем`3032. `эгекъэсхэ_ари_пч`3043. `ыгу,_цинащырыхэ_`305 306**Context Size 2:**307 3081. `гъэпсыр_зэрэ_ӏуад`3092. `ъэп_ву_адыгъэхьын`3103. `э_зыгэ_ж_дангьэ_т`311 312**Context Size 3:**313 3141. `гъэкъхэр,_кӏэ,_гум`3152. `_къэралыгъэдунэжъы`3163. `эм_и_–_зэрал_нэхэр`317 318**Context Size 4:**319 3201. `ыгъэ_гъэмрэ_приручи`3212. `хэр_бжъэдыгъуапэ_зэ`3223. `агъэхьан_хуейщ,_ахэ`323 324 325### Key Findings326 327- **Best Predictability:** Context-4 (word) with 98.7% predictability328- **Branching Factor:** Decreases with context size (more deterministic)329- **Memory Trade-off:** Larger contexts require more storage (73,859 contexts)330- **Recommendation:** Context-3 or Context-4 for text generation331 332---333## 4. Vocabulary Analysis334 335336 337338 339340 341### Statistics342 343| Metric | Value |344|--------|-------|345| Vocabulary Size | 7,120 |346| Total Tokens | 45,308 |347| Mean Frequency | 6.36 |348| Median Frequency | 3 |349| Frequency Std Dev | 22.08 |350 351### Most Common Words352 353| Rank | Word | Frequency |354|------|------|-----------|355| 1 | и | 999 |356| 2 | адыгэ | 660 |357| 3 | м | 508 |358| 4 | илъэсым | 406 |359| 5 | ащ | 391 |360| 6 | я | 320 |361| 7 | ары | 274 |362| 8 | а | 257 |363| 9 | нэбгырэ | 250 |364| 10 | е | 223 |365 366### Least Common Words (from vocabulary)367 368| Rank | Word | Frequency |369|------|------|-----------|370| 1 | muzea | 2 |371| 2 | britishpedia | 2 |372| 3 | encyklopedia | 2 |373| 4 | osobistości | 2 |374| 5 | rzeczypospolitej | 2 |375| 6 | polskiej | 2 |376| 7 | bph | 2 |377| 8 | british | 2 |378| 9 | publishing | 2 |379| 10 | ltd | 2 |380 381### Zipf's Law Analysis382 383| Metric | Value |384|--------|-------|385| Zipf Coefficient | 0.7863 |386| R² (Goodness of Fit) | 0.977814 |387| Adherence Quality | **excellent** |388 389### Coverage Analysis390 391| Top N Words | Coverage |392|-------------|----------|393| Top 100 | 28.8% |394| Top 1,000 | 60.5% |395| Top 5,000 | 90.6% |396| Top 10,000 | 0.0% |397 398### Key Findings399 400- **Zipf Compliance:** R²=0.9778 indicates excellent adherence to Zipf's law401- **High Frequency Dominance:** Top 100 words cover 28.8% of corpus402- **Long Tail:** -2,880 words needed for remaining 100.0% coverage403 404---405## 5. Word Embeddings Evaluation406 407408 409410 411412 413414 415 416### 5.1 Cross-Lingual Alignment417 418419 420421 422 423### 5.2 Model Comparison424 425| Model | Dimension | Isotropy | Semantic Density | Alignment R@1 | Alignment R@10 |426|-------|-----------|----------|------------------|---------------|----------------|427| **mono_32d** | 32 | 0.4880 | 0.4410 | N/A | N/A |428| **mono_64d** | 64 | 0.2186 | 0.3951 | N/A | N/A |429| **mono_128d** | 128 | 0.0372 | 0.3901 | N/A | N/A |430| **aligned_32d** | 32 | 0.4880 🏆 | 0.4477 | 0.0460 | 0.3851 |431| **aligned_64d** | 64 | 0.2186 | 0.3901 | 0.2011 | 0.7701 |432| **aligned_128d** | 128 | 0.0372 | 0.3927 | 0.2759 | 0.8103 |433 434### Key Findings435 436- **Best Isotropy:** aligned_32d with 0.4880 (more uniform distribution)437- **Semantic Density:** Average pairwise similarity of 0.4094. Lower values indicate better semantic separation.438- **Alignment Quality:** Aligned models achieve up to 27.6% R@1 in cross-lingual retrieval.439- **Recommendation:** 128d aligned for best cross-lingual performance440 441---442## 6. Morphological Analysis (Experimental)443 444This 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.445 446### 6.1 Productivity & Complexity447 448| Metric | Value | Interpretation | Recommendation |449|--------|-------|----------------|----------------|450| Productivity Index | **5.000** | High morphological productivity | Reliable analysis |451| Idiomaticity Gap | **0.610** | High formulaic/idiomatic content | - |452 453### 6.2 Affix Inventory (Productive Units)454 455These 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.456 457#### Productive Prefixes458| Prefix | Examples |459|--------|----------|460| `-къ` | къчр, къэлэшъо, къо |461| `-зэ` | зэрэхъугъэхэм, зэфэшъхьаф, зэхигъэуцогъэгъэ |462| `-къы` | къыӏуагъ, къыщыфэфедэщтхэу, къыгъэуцугъэ |463 464#### Productive Suffixes465| Suffix | Examples |466|--------|----------|467| `-э` | литературоведческэ, уиджыбэ, лъымрэ |468| `-м` | заповедникым, хъуагъэм, ипэм |469| `-р` | тхэныр, хунгариер, къчр |470| `-эр` | алъытэщтыгъэр, тхыбзэр, ылъэгъурэр |471| `-эм` | хъуагъэм, ипэм, псалъэжьхэм |472| `-эу` | цӏэу, дэлъэу, щысэу |473| `-хэр` | тыркухэр, ежьхэр, ахэр |474| `-рэ` | лъымрэ, цӏэмрэ, зыфиӏорэ |475 476### 6.3 Bound Stems (Lexical Roots)477 478Bound 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.479 480| Stem | Cohesion | Substitutability | Examples |481|------|----------|------------------|----------|482| `тыгъ` | 1.84x | 28 contexts | тыгъэ, тыгъу, итыгъ |483| `эпкъ` | 1.90x | 25 contexts | нэпкъ, тхэпкъ, лъэпкъ |484| `ъагъ` | 2.25x | 14 contexts | лъагъо, пчъагъ, пчъагъэ |485| `агъэ` | 1.63x | 39 contexts | благъэ, тхагъэ, пчагъэ |486| `дыгэ` | 2.03x | 14 contexts | адыгэ, адыгэу, адыгэм |487| `къуа` | 2.23x | 10 contexts | къуае, къуажэ, къуадж |488| `эхэр` | 1.72x | 20 contexts | бэхэр, дзэхэр, усэхэр |489| `ъхьэ` | 1.84x | 16 contexts | шъхьэ, пшъхьэ, шъхьэм |490| `псэу` | 1.70x | 20 contexts | упсэу, щэпсэу, щыпсэу |491| `шъхь` | 1.61x | 23 contexts | шъхьэ, пшъхьэ, шъхьэм |492| `ыгъо` | 1.66x | 19 contexts | цыгъо, мыгъо, цыгъор |493| `гъэх` | 1.79x | 14 contexts | багъэх, яӏагъэх, тхыгъэх |494 495### 6.4 Affix Compatibility (Co-occurrence)496 497This table shows which prefixes and suffixes most frequently co-occur on the same stems, revealing the 'stacking' rules of the language's morphology.498 499| Prefix | Suffix | Frequency | Examples |500|--------|--------|-----------|----------|501| `-къ` | `-э` | 94 words | къохьапӏэ, къыхаутыгъэ |502| `-къ` | `-р` | 64 words | къабзэр, къызэдыхэфэныр |503| `-къ` | `-м` | 56 words | къэралыгъуэм, къунетрэм |504| `-къ` | `-эр` | 52 words | къабзэр, къуаджэхэр |505| `-зэ` | `-р` | 43 words | зэреджэхэр, зэрар |506| `-зэ` | `-м` | 41 words | зэблэтхъуным, зэрагъэтэрэзыжьыгъэм |507| `-къ` | `-эм` | 36 words | къэралыгъуэм, къунетрэм |508| `-зэ` | `-эр` | 34 words | зэреджэхэр, зэпырыбгъэзэжьынхэр |509| `-къ` | `-эу` | 33 words | къыщегъэжьагъэу, къинэу |510| `-зэ` | `-э` | 31 words | зэкъотыныгъэ, зэралэжьырэ |511 512### 6.5 Recursive Morpheme Segmentation513 514Using **Recursive Hierarchical Substitutability**, we decompose complex words into their constituent morphemes. This approach handles nested affixes (e.g., `prefix-prefix-root-suffix`).515 516| Word | Suggested Split | Confidence | Stem |517|------|-----------------|------------|------|518| республикэмрэ | **`республик-эм-рэ`** | 6.0 | `республик` |519| макъэхэмрэ | **`макъэ-хэм-рэ`** | 6.0 | `макъэ` |520| литературэмрэ | **`литератур-эм-рэ`** | 6.0 | `литератур` |521| благъохэмрэ | **`благъо-хэм-рэ`** | 6.0 | `благъо` |522| бзылъфыгъэмрэ | **`бзылъфыгъ-эм-рэ`** | 6.0 | `бзылъфыгъ` |523| литературэр | **`литератур-эр`** | 4.5 | `литератур` |524| диалектэу | **`диалект-эу`** | 4.5 | `диалект` |525| агъэфедэрэ | **`агъэфедэ-рэ`** | 4.5 | `агъэфедэ` |526| шъхьафитэу | **`шъхьафит-эу`** | 4.5 | `шъхьафит` |527| зыкъэзыӏэтыгъэм | **`зыкъэзыӏэтыгъ-эм`** | 4.5 | `зыкъэзыӏэтыгъ` |528| ишъхъэрэмрэ | **`ишъхъ-эр-эм-рэ`** | 4.5 | `ишъхъ` |529| адэмыехэр | **`адэмые-хэр`** | 4.5 | `адэмые` |530| зэкъоуцохэу | **`зэ-къ-оуцох-эу`** | 4.5 | `оуцох` |531| ыгузэгухэм | **`ыгузэгу-хэм`** | 4.5 | `ыгузэгу` |532| беслъэнейхэр | **`беслъэней-хэр`** | 4.5 | `беслъэней` |533 534### 6.6 Linguistic Interpretation535 536> **Automated Insight:**537The language Adyghe shows high morphological productivity. The subword models are significantly more efficient than word models, suggesting a rich system of affixation or compounding.538 539> **Note on Idiomaticity:** The high Idiomaticity Gap suggests a large number of frequent multi-word expressions or formulaic sequences that are statistically distinct from their component parts.540 541---542## 7. Summary & Recommendations543 544545 546### Production Recommendations547 548| Component | Recommended | Rationale |549|-----------|-------------|-----------|550| Tokenizer | **32k BPE** | Best compression (4.20x) |551| N-gram | **2-gram** | Lowest perplexity (407) |552| Markov | **Context-4** | Highest predictability (98.7%) |553| Embeddings | **100d** | Balanced semantic capture and isotropy |554 555 556---557## Appendix: Metrics Glossary & Interpretation Guide558 559This section provides definitions, intuitions, and guidance for interpreting the metrics used throughout this report.560 561### Tokenizer Metrics562 563**Compression Ratio**564> *Definition:* The ratio of characters to tokens (chars/token). Measures how efficiently the tokenizer represents text.565>566> *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.567>568> *What to seek:* Higher is generally better for efficiency, but extremely high compression may indicate overly aggressive merging that loses morphological information.569 570**Average Token Length (Fertility)**571> *Definition:* Mean number of characters per token produced by the tokenizer.572>573> *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.574>575> *What to seek:* Balance between 2-5 characters for most languages. Arabic/morphologically-rich languages may benefit from slightly longer tokens.576 577**Unknown Token Rate (OOV Rate)**578> *Definition:* Percentage of tokens that map to the unknown/UNK token, indicating words the tokenizer cannot represent.579>580> *Intuition:* Lower OOV means better vocabulary coverage. High OOV indicates the tokenizer encounters many unseen character sequences.581>582> *What to seek:* Below 1% is excellent; below 5% is acceptable. BPE tokenizers typically achieve very low OOV due to subword fallback.583 584### N-gram Model Metrics585 586**Perplexity**587> *Definition:* Measures how "surprised" the model is by test data. Mathematically: 2^(cross-entropy). Lower values indicate better prediction.588>589> *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.590>591> *What to seek:* Lower is better. Perplexity decreases with larger n-grams (more context). Values vary widely by language and corpus size.592 593**Entropy**594> *Definition:* Average information content (in bits) needed to encode the next token given the context. Related to perplexity: perplexity = 2^entropy.595>596> *Intuition:* High entropy means high uncertainty/randomness; low entropy means predictable patterns. Natural language typically has entropy between 1-4 bits per character.597>598> *What to seek:* Lower entropy indicates more predictable text patterns. Entropy should decrease as n-gram size increases.599 600**Coverage (Top-K)**601> *Definition:* Percentage of corpus occurrences explained by the top K most frequent n-grams.602>603> *Intuition:* High coverage with few patterns indicates repetitive/formulaic text; low coverage suggests diverse vocabulary usage.604>605> *What to seek:* Depends on use case. For language modeling, moderate coverage (40-60% with top-1000) is typical for natural text.606 607### Markov Chain Metrics608 609**Average Entropy**610> *Definition:* Mean entropy across all contexts, measuring average uncertainty in next-word prediction.611>612> *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).613>614> *What to seek:* Decreasing entropy with larger context sizes. Very low entropy (<0.1) indicates highly deterministic transitions.615 616**Branching Factor**617> *Definition:* Average number of unique next tokens observed for each context.618>619> *Intuition:* High branching = many possible continuations (flexible but uncertain); low branching = few options (predictable but potentially repetitive).620>621> *What to seek:* Branching factor should decrease with context size. Values near 1.0 indicate nearly deterministic chains.622 623**Predictability**624> *Definition:* Derived metric: (1 - normalized_entropy) × 100%. Indicates how deterministic the model's predictions are.625>626> *Intuition:* 100% predictability means the next word is always certain; 0% means completely random. Real text falls between these extremes.627>628> *What to seek:* Higher predictability for text generation quality, but too high (>98%) may produce repetitive output.629 630### Vocabulary & Zipf's Law Metrics631 632**Zipf's Coefficient**633> *Definition:* The slope of the log-log plot of word frequency vs. rank. Zipf's law predicts this should be approximately -1.634>635> *Intuition:* A coefficient near -1 indicates the corpus follows natural language patterns where a few words are very common and most words are rare.636>637> *What to seek:* Values between -0.8 and -1.2 indicate healthy natural language distribution. Deviations may suggest domain-specific or artificial text.638 639**R² (Coefficient of Determination)**640> *Definition:* Measures how well the linear fit explains the frequency-rank relationship. Ranges from 0 to 1.641>642> *Intuition:* R² near 1.0 means the data closely follows Zipf's law; lower values indicate deviation from expected word frequency patterns.643>644> *What to seek:* R² > 0.95 is excellent; > 0.99 indicates near-perfect Zipf adherence typical of large natural corpora.645 646**Vocabulary Coverage**647> *Definition:* Cumulative percentage of corpus tokens accounted for by the top N words.648>649> *Intuition:* Shows how concentrated word usage is. If top-100 words cover 50% of text, the corpus relies heavily on common words.650>651> *What to seek:* Top-100 covering 30-50% is typical. Higher coverage indicates more repetitive text; lower suggests richer vocabulary.652 653### Word Embedding Metrics654 655**Isotropy**656> *Definition:* Measures how uniformly distributed vectors are in the embedding space. Computed as the ratio of minimum to maximum singular values.657>658> *Intuition:* High isotropy (near 1.0) means vectors spread evenly in all directions; low isotropy means vectors cluster in certain directions, reducing expressiveness.659>660> *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.661 662**Average Norm**663> *Definition:* Mean magnitude (L2 norm) of word vectors in the embedding space.664>665> *Intuition:* Indicates the typical "length" of vectors. Consistent norms suggest stable training; high variance may indicate some words are undertrained.666>667> *What to seek:* Relatively consistent norms across models. The absolute value matters less than consistency (low std deviation).668 669**Cosine Similarity**670> *Definition:* Measures angular similarity between vectors, ranging from -1 (opposite) to 1 (identical direction).671>672> *Intuition:* Words with similar meanings should have high cosine similarity. This is the standard metric for semantic relatedness in embeddings.673>674> *What to seek:* Semantically related words should score > 0.5; unrelated words should be near 0. Synonyms often score > 0.7.675 676**t-SNE Visualization**677> *Definition:* t-Distributed Stochastic Neighbor Embedding - a dimensionality reduction technique that preserves local structure for visualization.678>679> *Intuition:* Clusters in t-SNE plots indicate groups of semantically related words. Spread indicates vocabulary diversity; tight clusters suggest semantic coherence.680>681> *What to seek:* Meaningful clusters (e.g., numbers together, verbs together). Avoid over-interpreting distances - t-SNE preserves local, not global, structure.682 683### General Interpretation Guidelines684 6851. **Compare within model families:** Metrics are most meaningful when comparing models of the same type (e.g., 8k vs 64k tokenizer).6862. **Consider trade-offs:** Better performance on one metric often comes at the cost of another (e.g., compression vs. OOV rate).6873. **Context matters:** Optimal values depend on downstream tasks. Text generation may prioritize different metrics than classification.6884. **Corpus influence:** All metrics are influenced by corpus characteristics. Wikipedia text differs from social media or literature.6895. **Language-specific patterns:** Morphologically rich languages (like Arabic) may show different optimal ranges than analytic languages.690 691 692### Visualizations Index693 694| Visualization | Description |695|---------------|-------------|696| Tokenizer Compression | Compression ratios by vocabulary size |697| Tokenizer Fertility | Average token length by vocabulary |698| Tokenizer OOV | Unknown token rates |699| Tokenizer Total Tokens | Total tokens by vocabulary |700| N-gram Perplexity | Perplexity by n-gram size |701| N-gram Entropy | Entropy by n-gram size |702| N-gram Coverage | Top pattern coverage |703| N-gram Unique | Unique n-gram counts |704| Markov Entropy | Entropy by context size |705| Markov Branching | Branching factor by context |706| Markov Contexts | Unique context counts |707| Zipf's Law | Frequency-rank distribution with fit |708| Vocab Frequency | Word frequency distribution |709| Top 20 Words | Most frequent words |710| Vocab Coverage | Cumulative coverage curve |711| Embedding Isotropy | Vector space uniformity |712| Embedding Norms | Vector magnitude distribution |713| Embedding Similarity | Word similarity heatmap |714| Nearest Neighbors | Similar words for key terms |715| t-SNE Words | 2D word embedding visualization |716| t-SNE Sentences | 2D sentence embedding visualization |717| Position Encoding | Encoding method comparison |718| Model Sizes | Storage requirements |719| Performance Dashboard | Comprehensive performance overview |720 721---722## About This Project723 724### Data Source725 726Models trained on [wikipedia-monthly](https://huggingface.co/datasets/omarkamali/wikipedia-monthly) - a monthly snapshot of Wikipedia articles across 300+ languages.727 728### Project729 730A project by **[Wikilangs](https://wikilangs.org)** - Open-source NLP models for every Wikipedia language.731 732### Maintainer733 734[Omar Kamali](https://omarkamali.com) - [Omneity Labs](https://omneitylabs.com)735 736### Citation737 738If you use these models in your research, please cite:739 740```bibtex741@misc{wikilangs2025,742 author = {Kamali, Omar},743 title = {Wikilangs: Open NLP Models for Wikipedia Languages},744 year = {2025},745 doi = {10.5281/zenodo.18073153},746 publisher = {Zenodo},747 url = {https://huggingface.co/wikilangs}748 institution = {Omneity Labs}749}750```751 752### License753 754MIT License - Free for academic and commercial use.755 756### Links757 758- 🌐 Website: [wikilangs.org](https://wikilangs.org)759- 🤗 Models: [huggingface.co/wikilangs](https://huggingface.co/wikilangs)760- 📊 Data: [wikipedia-monthly](https://huggingface.co/datasets/omarkamali/wikipedia-monthly)761- 👤 Author: [Omar Kamali](https://huggingface.co/omarkamali)762- 🤝 Sponsor: [Featherless AI](https://featherless.ai)763---764*Generated by Wikilangs Models Pipeline*765 766*Report Date: 2026-01-03 18:25:02*767