omarkamali commited on
Commit
abfac1f
·
verified ·
1 Parent(s): dcc0ae9

Upload all models and assets for az (latest)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. README.md +214 -179
  3. models/embeddings/aligned/az_128d.bin +3 -0
  4. models/embeddings/aligned/az_128d.meta.json +1 -0
  5. models/embeddings/aligned/az_128d.projection.npy +3 -0
  6. models/embeddings/aligned/az_128d_metadata.json +8 -0
  7. models/embeddings/aligned/az_32d.bin +3 -0
  8. models/embeddings/aligned/az_32d.meta.json +1 -0
  9. models/embeddings/aligned/az_32d.projection.npy +3 -0
  10. models/embeddings/aligned/az_32d_metadata.json +8 -0
  11. models/embeddings/aligned/az_64d.bin +3 -0
  12. models/embeddings/aligned/az_64d.meta.json +1 -0
  13. models/embeddings/aligned/az_64d.projection.npy +3 -0
  14. models/embeddings/aligned/az_64d_metadata.json +8 -0
  15. models/embeddings/monolingual/az_128d.bin +2 -2
  16. models/embeddings/monolingual/az_128d_metadata.json +1 -1
  17. models/embeddings/monolingual/az_32d.bin +2 -2
  18. models/embeddings/monolingual/az_32d_metadata.json +1 -1
  19. models/embeddings/monolingual/az_64d.bin +2 -2
  20. models/embeddings/monolingual/az_64d_metadata.json +1 -1
  21. models/subword_markov/az_markov_ctx1_subword.parquet +2 -2
  22. models/subword_markov/az_markov_ctx1_subword_metadata.json +2 -2
  23. models/subword_markov/az_markov_ctx2_subword.parquet +2 -2
  24. models/subword_markov/az_markov_ctx2_subword_metadata.json +2 -2
  25. models/subword_markov/az_markov_ctx3_subword.parquet +2 -2
  26. models/subword_markov/az_markov_ctx3_subword_metadata.json +2 -2
  27. models/subword_markov/az_markov_ctx4_subword.parquet +2 -2
  28. models/subword_markov/az_markov_ctx4_subword_metadata.json +2 -2
  29. models/subword_ngram/az_2gram_subword.parquet +2 -2
  30. models/subword_ngram/az_2gram_subword_metadata.json +2 -2
  31. models/subword_ngram/az_3gram_subword.parquet +2 -2
  32. models/subword_ngram/az_3gram_subword_metadata.json +2 -2
  33. models/subword_ngram/az_4gram_subword.parquet +2 -2
  34. models/subword_ngram/az_4gram_subword_metadata.json +2 -2
  35. models/subword_ngram/az_5gram_subword.parquet +3 -0
  36. models/subword_ngram/az_5gram_subword_metadata.json +7 -0
  37. models/tokenizer/az_tokenizer_16k.model +2 -2
  38. models/tokenizer/az_tokenizer_16k.vocab +0 -0
  39. models/tokenizer/az_tokenizer_32k.model +2 -2
  40. models/tokenizer/az_tokenizer_32k.vocab +0 -0
  41. models/tokenizer/az_tokenizer_64k.model +2 -2
  42. models/tokenizer/az_tokenizer_64k.vocab +0 -0
  43. models/tokenizer/az_tokenizer_8k.model +2 -2
  44. models/tokenizer/az_tokenizer_8k.vocab +0 -0
  45. models/vocabulary/az_vocabulary.parquet +2 -2
  46. models/vocabulary/az_vocabulary_metadata.json +9 -9
  47. models/word_markov/az_markov_ctx1_word.parquet +2 -2
  48. models/word_markov/az_markov_ctx1_word_metadata.json +2 -2
  49. models/word_markov/az_markov_ctx2_word.parquet +2 -2
  50. models/word_markov/az_markov_ctx2_word_metadata.json +2 -2
.gitattributes CHANGED
@@ -39,3 +39,4 @@ visualizations/position_encoding_comparison.png filter=lfs diff=lfs merge=lfs -t
39
  visualizations/tsne_sentences.png filter=lfs diff=lfs merge=lfs -text
40
  visualizations/tsne_words.png filter=lfs diff=lfs merge=lfs -text
41
  visualizations/zipf_law.png filter=lfs diff=lfs merge=lfs -text
 
 
39
  visualizations/tsne_sentences.png filter=lfs diff=lfs merge=lfs -text
40
  visualizations/tsne_words.png filter=lfs diff=lfs merge=lfs -text
41
  visualizations/zipf_law.png filter=lfs diff=lfs merge=lfs -text
42
+ visualizations/embedding_tsne_multilingual.png filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,6 +1,6 @@
1
  ---
2
  language: az
3
- language_name: AZ
4
  language_family: turkic_oghuz
5
  tags:
6
  - wikilangs
@@ -10,11 +10,21 @@ tags:
10
  - n-gram
11
  - markov
12
  - wikipedia
 
 
 
 
 
 
 
 
 
 
13
  - monolingual
14
  - family-turkic_oghuz
15
  license: mit
16
  library_name: wikilangs
17
- pipeline_tag: feature-extraction
18
  datasets:
19
  - omarkamali/wikipedia-monthly
20
  dataset_info:
@@ -23,20 +33,20 @@ dataset_info:
23
  metrics:
24
  - name: best_compression_ratio
25
  type: compression
26
- value: 5.127
27
  - name: best_isotropy
28
  type: isotropy
29
- value: 0.8147
30
  - name: vocabulary_size
31
  type: vocab
32
  value: 0
33
- generated: 2026-01-03
34
  ---
35
 
36
- # AZ - Wikilangs Models
37
  ## Comprehensive Research Report & Full Ablation Study
38
 
39
- This repository contains NLP models trained and evaluated by Wikilangs, specifically on **AZ** Wikipedia data.
40
  We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and word embeddings.
41
 
42
  ## 📋 Repository Contents
@@ -60,7 +70,7 @@ We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and
60
  - [3. Markov Chain Evaluation](#3-markov-chain-evaluation)
61
  - [4. Vocabulary Analysis](#4-vocabulary-analysis)
62
  - [5. Word Embeddings Evaluation](#5-word-embeddings-evaluation)
63
- - [6. Morphological Analysis (Experimental)](#6-morphological-analysis)
64
  - [7. Summary & Recommendations](#7-summary--recommendations)
65
  - [Metrics Glossary](#appendix-metrics-glossary--interpretation-guide)
66
  - [Visualizations Index](#visualizations-index)
@@ -80,25 +90,16 @@ We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and
80
 
81
  | Vocab Size | Compression | Avg Token Len | UNK Rate | Total Tokens |
82
  |------------|-------------|---------------|----------|--------------|
83
- | **8k** | 3.940x | 3.94 | 0.0953% | 1,262,968 |
84
- | **16k** | 4.420x | 4.42 | 0.1069% | 1,125,834 |
85
- | **32k** | 4.818x | 4.82 | 0.1165% | 1,032,837 |
86
- | **64k** | 5.127x 🏆 | 5.13 | 0.1239% | 970,666 |
87
 
88
  ### Tokenization Examples
89
 
90
  Below are sample sentences tokenized with each vocabulary size:
91
 
92
- **Sample 1:** `Bitlis vilayəti Osmanlı İmperiyası tərkibində, illərdə mövcud olmuş I dərəcəli...`
93
-
94
- | Vocab | Tokens | Count |
95
- |-------|--------|-------|
96
- | 8k | `▁bit lis ▁vilayəti ▁— ▁osmanlı ▁İmper iyası ▁tərkibində , ▁illərdə ... (+17 more)` | 27 |
97
- | 16k | `▁bit lis ▁vilayəti ▁— ▁osmanlı ▁İmperiyası ▁tərkibində , ▁illərdə ▁mövcud ... (+16 more)` | 26 |
98
- | 32k | `▁bit lis ▁vilayəti ▁— ▁osmanlı ▁İmperiyası ▁tərkibində , ▁illərdə ▁mövcud ... (+16 more)` | 26 |
99
- | 64k | `▁bitlis ▁vilayəti ▁— ▁osmanlı ▁İmperiyası ▁tərkibində , ▁illərdə ▁mövcud ▁olmuş ... (+14 more)` | 24 |
100
-
101
- **Sample 2:** `() — aləminin dəstəsinin fəsiləsinə aid bitki cinsi. Sinonimləri Heterotipik sin...`
102
 
103
  | Vocab | Tokens | Count |
104
  |-------|--------|-------|
@@ -107,20 +108,29 @@ Below are sample sentences tokenized with each vocabulary size:
107
  | 32k | `▁() ▁— ▁aləminin ▁dəstəsinin ▁fəsiləsinə ▁aid ▁bitki ▁cinsi . ▁sinonimləri ... (+6 more)` | 16 |
108
  | 64k | `▁() ▁— ▁aləminin ▁dəstəsinin ▁fəsiləsinə ▁aid ▁bitki ▁cinsi . ▁sinonimləri ... (+6 more)` | 16 |
109
 
110
- **Sample 3:** `Üçüncü simfoniya (film, Üçüncü simfoniya (Motsart) Üçüncü simfoniya (Çaykovski) ...`
111
 
112
  | Vocab | Tokens | Count |
113
  |-------|--------|-------|
114
- | 8k | `▁üçüncüsimf oniya( film , ▁üçüncüsimf oniya ▁( ... (+21 more)` | 31 |
115
- | 16k | `▁üçüncüsimf oniya( film , ▁üçüncüsimf oniya ▁( ... (+17 more)` | 27 |
116
- | 32k | `▁üçüncüsimfoniya( film , ▁üçüncüsimfoniya( mot sart ... (+13 more)` | 23 |
117
- | 64k | `▁üçüncüsimfoniya( film , ▁üçüncüsimfoniya( mot sart ... (+13 more)` | 23 |
 
 
 
 
 
 
 
 
 
118
 
119
 
120
  ### Key Findings
121
 
122
- - **Best Compression:** 64k achieves 5.127x compression
123
- - **Lowest UNK Rate:** 8k with 0.0953% unknown tokens
124
  - **Trade-off:** Larger vocabularies improve compression but increase model size
125
  - **Recommendation:** 32k vocabulary provides optimal balance for production use
126
 
@@ -137,12 +147,14 @@ Below are sample sentences tokenized with each vocabulary size:
137
 
138
  | N-gram | Variant | Perplexity | Entropy | Unique N-grams | Top-100 Coverage | Top-1000 Coverage |
139
  |--------|---------|------------|---------|----------------|------------------|-------------------|
140
- | **2-gram** | Word | 266,647 | 18.02 | 1,217,523 | 4.8% | 13.7% |
141
- | **2-gram** | Subword | 405 🏆 | 8.66 | 18,177 | 58.1% | 97.7% |
142
- | **3-gram** | Word | 580,086 | 19.15 | 1,735,864 | 4.1% | 9.8% |
143
- | **3-gram** | Subword | 3,752 | 11.87 | 159,097 | 20.7% | 61.1% |
144
- | **4-gram** | Word | 1,224,902 | 20.22 | 3,019,123 | 3.9% | 8.4% |
145
- | **4-gram** | Subword | 21,204 | 14.37 | 964,243 | 10.3% | 32.7% |
 
 
146
 
147
  ### Top 5 N-grams by Size
148
 
@@ -150,68 +162,88 @@ Below are sample sentences tokenized with each vocabulary size:
150
 
151
  | Rank | N-gram | Count |
152
  |------|--------|-------|
153
- | 1 | `və ya` | 82,539 |
154
- | 2 | `xarici keçidlər` | 64,635 |
155
- | 3 | `həmçinin bax` | 61,223 |
156
- | 4 | `i̇stinadlar xarici` | 45,022 |
157
- | 5 | `i̇stinadlar həmçinin` | 30,535 |
158
 
159
  **3-grams (Word):**
160
 
161
  | Rank | N-gram | Count |
162
  |------|--------|-------|
163
- | 1 | `i̇stinadlar xarici keçidlər` | 44,533 |
164
- | 2 | `i̇stinadlar həmçinin bax` | 30,508 |
165
- | 3 | `fəsiləsinin cinsinə aid` | 20,829 |
166
- | 4 | `dəstəsinin fəsiləsinin cinsinə` | 18,108 |
167
- | 5 | `aid bitki növü` | 17,478 |
168
 
169
  **4-grams (Word):**
170
 
171
  | Rank | N-gram | Count |
172
  |------|--------|-------|
173
- | 1 | `dəstəsinin fəsiləsinin cinsinə aid` | 18,108 |
174
- | 2 | `cinsinə aid bitki növü` | 17,459 |
175
- | 3 | `fəsiləsinin cinsinə aid bitki` | 17,424 |
176
- | 4 | `aləminin dəstəsinin fəsiləsinin cinsinə` | 14,444 |
177
- | 5 | `növü i̇stinadlar həmçinin bax` | 10,412 |
 
 
 
 
 
 
 
 
 
 
178
 
179
  **2-grams (Subword):**
180
 
181
  | Rank | N-gram | Count |
182
  |------|--------|-------|
183
- | 1 | `n _` | 7,988,809 |
184
- | 2 | `ə _` | 6,442,941 |
185
- | 3 | `i n` | 6,166,226 |
186
- | 4 | `a r` | 5,329,650 |
187
- | 5 | `ə r` | 5,265,570 |
188
 
189
  **3-grams (Subword):**
190
 
191
  | Rank | N-gram | Count |
192
  |------|--------|-------|
193
- | 1 | `l ə r` | 2,404,591 |
194
- | 2 | `l a r` | 2,255,189 |
195
- | 3 | `d ə _` | 2,141,762 |
196
- | 4 | `i n _` | 2,027,629 |
197
- | 5 | `a n _` | 1,821,260 |
198
 
199
  **4-grams (Subword):**
200
 
201
  | Rank | N-gram | Count |
202
  |------|--------|-------|
203
- | 1 | `_ v ə _` | 1,462,860 |
204
- | 2 | `l ə r i` | 1,237,145 |
205
- | 3 | `l a r ı` | 1,052,371 |
206
- | 4 | `i n d ə` | 1,049,474 |
207
- | 5 | `n i n _` | 951,332 |
 
 
 
 
 
 
 
 
 
 
208
 
209
 
210
  ### Key Findings
211
 
212
- - **Best Perplexity:** 2-gram (subword) with 405
213
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
214
- - **Coverage:** Top-1000 patterns cover ~33% of corpus
215
  - **Recommendation:** 4-gram or 5-gram for best predictive performance
216
 
217
  ---
@@ -227,14 +259,14 @@ Below are sample sentences tokenized with each vocabulary size:
227
 
228
  | Context | Variant | Avg Entropy | Perplexity | Branching Factor | Unique Contexts | Predictability |
229
  |---------|---------|-------------|------------|------------------|-----------------|----------------|
230
- | **1** | Word | 0.9394 | 1.918 | 11.41 | 1,714,220 | 6.1% |
231
- | **1** | Subword | 1.1697 | 2.250 | 8.00 | 8,084 | 0.0% |
232
- | **2** | Word | 0.3187 | 1.247 | 1.95 | 19,534,498 | 68.1% |
233
- | **2** | Subword | 0.7478 | 1.679 | 5.29 | 64,659 | 25.2% |
234
- | **3** | Word | 0.1043 | 1.075 | 1.20 | 37,988,863 | 89.6% |
235
- | **3** | Subword | 0.8132 | 1.757 | 4.77 | 341,910 | 18.7% |
236
- | **4** | Word | 0.0351 🏆 | 1.025 | 1.05 | 45,491,212 | 96.5% |
237
- | **4** | Subword | 0.7291 | 1.658 | 3.64 | 1,630,545 | 27.1% |
238
 
239
  ### Generated Text Samples (Word-based)
240
 
@@ -242,27 +274,27 @@ Below are text samples generated from each word-based Markov chain model:
242
 
243
  **Context Size 1:**
244
 
245
- 1. `və digər qarağac küçəsində yerləşən bike travel channel saytında volodya rubinin hekayəsinin sadələş...`
246
- 2. `ildə göyçə kanalı 2 4 kimi qəbul ya radiodalğaların ortaya çıxmışdı fransız şərqşünas vital bakım`
247
- 3. `ilə rəqabət üstünlükləri üçün ilham alınmışdır yava nın rəhbəri aqrar sahədə süni tıxacı qanunu pozm...`
248
 
249
  **Context Size 2:**
250
 
251
- 1. `və ya onun mənzilinin qədər qorxunc idilər həmin gün masovkada iştirak etməyi özləri üçün nisbətə...`
252
- 2. `xarici keçidlər hissi`
253
- 3. `i̇stinadlar xarici keçidlər romario sambafoot com romario siyasətə qatıldı futbolçuları fk oyunçular...`
254
 
255
  **Context Size 3:**
256
 
257
- 1. `i̇stinadlar xarici keçidlər середа с а перспективы охраны авторских и смежных прав в условиях распро...`
258
- 2. `fəsiləsinin cinsinə aid heyvan növü i̇stinadlar həmçinin bax aprel işğalı əlavə ədəbiyyat sovet sosi...`
259
- 3. `dəstəsinin fəsiləsinin cinsinə aid bitki növü i̇stinadlar həmçinin bax koreyanın xüsusi şəhərləri i̇...`
260
 
261
  **Context Size 4:**
262
 
263
- 1. `dəstəsinin fəsiləsinin cinsinə aid heyvan növü i̇stinadlar həmçinin bax ildə təsvir edilən bitkilər ...`
264
- 2. `cinsinə aid bitki növü i̇stinadlar həmçinin bax ildə təsvir edilən bitkilər linney tərəfindən adland...`
265
- 3. `fəsiləsinin cinsinə aid bitki növü ulvanın tallomu lövhəşəkilli parlaq yaşıl rəngli olub kənarları b...`
266
 
267
 
268
  ### Generated Text Samples (Subword-based)
@@ -271,34 +303,34 @@ Below are text samples generated from each subword-based Markov chain model:
271
 
272
  **Context Size 1:**
273
 
274
- 1. `_dalmuny_isuməla`
275
- 2. `allədi,_ilqəbulə`
276
- 3. `i_1._sın_lirpriy`
277
 
278
  **Context Size 2:**
279
 
280
- 1. `n_ya_tabı_vəfalar`
281
- 2. `ə_birilarabaxi:_а`
282
- 3. `inəşdira_meyvali_`
283
 
284
  **Context Size 3:**
285
 
286
- 1. `ləri_və_şimalınmas`
287
- 2. `ları_ekspilm)_+рас`
288
- 3. `də_onlar,_ərbi_tağ`
289
 
290
  **Context Size 4:**
291
 
292
- 1. `_və_kəndləri_kimi_f`
293
- 2. `ləri,_ildə_etdirir.`
294
- 3. `ində_aztv_“günorta_`
295
 
296
 
297
  ### Key Findings
298
 
299
  - **Best Predictability:** Context-4 (word) with 96.5% predictability
300
  - **Branching Factor:** Decreases with context size (more deterministic)
301
- - **Memory Trade-off:** Larger contexts require more storage (1,630,545 contexts)
302
  - **Recommendation:** Context-3 or Context-4 for text generation
303
 
304
  ---
@@ -314,64 +346,64 @@ Below are text samples generated from each subword-based Markov chain model:
314
 
315
  | Metric | Value |
316
  |--------|-------|
317
- | Vocabulary Size | 753,394 |
318
- | Total Tokens | 53,281,406 |
319
- | Mean Frequency | 70.72 |
320
  | Median Frequency | 4 |
321
- | Frequency Std Dev | 2273.78 |
322
 
323
  ### Most Common Words
324
 
325
  | Rank | Word | Frequency |
326
  |------|------|-----------|
327
- | 1 | və | 1,467,833 |
328
- | 2 | ildə | 410,483 |
329
- | 3 | ilə | 408,441 |
330
- | 4 | bir | 361,923 |
331
- | 5 | bu | 355,796 |
332
- | 6 | də | 228,505 |
333
- | 7 | azərbaycan | 220,507 |
334
- | 8 | üçün | 219,884 |
335
- | 9 | olan | 219,730 |
336
- | 10 | sonra | 179,588 |
337
 
338
  ### Least Common Words (from vocabulary)
339
 
340
  | Rank | Word | Frequency |
341
  |------|------|-----------|
342
- | 1 | i̇netimi | 2 |
343
- | 2 | timayanın | 2 |
344
- | 3 | llnp | 2 |
345
- | 4 | moqrovexonun | 2 |
346
- | 5 | məhkəməsiazərbaycan | 2 |
347
- | 6 | nəbiqə | 2 |
348
- | 7 | zübyani | 2 |
349
- | 8 | əşanı | 2 |
350
- | 9 | tülücü | 2 |
351
- | 10 | yenidoğulanlar | 2 |
352
 
353
  ### Zipf's Law Analysis
354
 
355
  | Metric | Value |
356
  |--------|-------|
357
  | Zipf Coefficient | 0.9645 |
358
- | R² (Goodness of Fit) | 0.992332 |
359
  | Adherence Quality | **excellent** |
360
 
361
  ### Coverage Analysis
362
 
363
  | Top N Words | Coverage |
364
  |-------------|----------|
365
- | Top 100 | 20.7% |
366
  | Top 1,000 | 45.3% |
367
- | Top 5,000 | 65.4% |
368
  | Top 10,000 | 73.7% |
369
 
370
  ### Key Findings
371
 
372
- - **Zipf Compliance:** R²=0.9923 indicates excellent adherence to Zipf's law
373
- - **High Frequency Dominance:** Top 100 words cover 20.7% of corpus
374
- - **Long Tail:** 743,394 words needed for remaining 26.3% coverage
375
 
376
  ---
377
  ## 5. Word Embeddings Evaluation
@@ -387,37 +419,40 @@ Below are text samples generated from each subword-based Markov chain model:
387
 
388
  ### 5.1 Cross-Lingual Alignment
389
 
390
- > *Note: Multilingual alignment visualization not available for this language.*
 
 
391
 
392
 
393
  ### 5.2 Model Comparison
394
 
395
  | Model | Dimension | Isotropy | Semantic Density | Alignment R@1 | Alignment R@10 |
396
  |-------|-----------|----------|------------------|---------------|----------------|
397
- | **mono_32d** | 32 | 0.8147 🏆 | 0.3523 | N/A | N/A |
398
- | **mono_64d** | 64 | 0.8067 | 0.2814 | N/A | N/A |
399
- | **mono_128d** | 128 | 0.7697 | 0.2228 | N/A | N/A |
 
 
 
400
 
401
  ### Key Findings
402
 
403
- - **Best Isotropy:** mono_32d with 0.8147 (more uniform distribution)
404
- - **Semantic Density:** Average pairwise similarity of 0.2855. Lower values indicate better semantic separation.
405
- - **Alignment Quality:** No aligned models evaluated in this run.
406
  - **Recommendation:** 128d aligned for best cross-lingual performance
407
 
408
  ---
409
  ## 6. Morphological Analysis (Experimental)
410
 
411
- > ⚠️ **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.
412
-
413
  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.
414
 
415
  ### 6.1 Productivity & Complexity
416
 
417
  | Metric | Value | Interpretation | Recommendation |
418
  |--------|-------|----------------|----------------|
419
- | Productivity Index | **0.000** | Low morphological productivity | ⚠️ Likely unreliable |
420
- | Idiomaticity Gap | **-1.000** | Low formulaic content | - |
421
 
422
  ### 6.2 Affix Inventory (Productive Units)
423
 
@@ -430,14 +465,14 @@ These are the most productive prefixes and suffixes identified by sampling the v
430
  #### Productive Suffixes
431
  | Suffix | Examples |
432
  |--------|----------|
433
- | `-n` | qanqlionlarından, kreyin, i̇radənin |
434
- | `-a` | aoyama, puraka, şahhüseynova |
435
- | `-in` | kreyin, i̇radənin, naimin |
436
- | `-an` | qanqlionlarından, saldıran, məmulatdan |
437
- | `-ın` | hanedanının, quldurlarının, çarımın |
438
- | `-dan` | qanqlionlarından, məmulatdan, yazıçılarından |
439
- | `-ən` | sevgiən, kombateldən, filmdəkindən |
440
- | `-nın` | hanedanının, quldurlarının, medyanın |
441
 
442
  ### 6.3 Bound Stems (Lexical Roots)
443
 
@@ -445,18 +480,18 @@ Bound stems are high-frequency subword units that are semantically cohesive but
445
 
446
  | Stem | Cohesion | Substitutability | Examples |
447
  |------|----------|------------------|----------|
448
- | `ərba` | 2.77x | 42 contexts | ərbaş, bərba, ərbab |
449
- | `ayca` | 2.92x | 24 contexts | cayca, aycan, sayca |
450
- | `rbay` | 2.33x | 54 contexts | erbay, arbay, orbay |
451
- | `arix` | 2.02x | 69 contexts | tarix, farix, larix |
452
- | `nlar` | 1.37x | 429 contexts | anlar, onlar, nları |
453
- | `irlə` | 1.36x | 390 contexts | pirlə, birlə, virlə |
454
- | `mişd` | 1.58x | 164 contexts | mişdi, emişdi, mişdir |
455
- | `mışd` | 1.62x | 142 contexts | mışdı, mışdır, aşmışdı |
456
- | `rləş` | 1.82x | 76 contexts | yerləş, birləş, yrləşən |
457
- | `ycan` | 2.87x | 13 contexts | aycan, beycan, bəycan |
458
- | `ərəf` | 1.65x | 87 contexts | ərəfə, tərəf, şərəf |
459
- | `qlar` | 1.38x | 199 contexts | aqlar, qlarn, doqlar |
460
 
461
  ### 6.4 Affix Compatibility (Co-occurrence)
462
 
@@ -471,26 +506,26 @@ Using **Recursive Hierarchical Substitutability**, we decompose complex words in
471
 
472
  | Word | Suggested Split | Confidence | Stem |
473
  |------|-----------------|------------|------|
474
- | şəyirdlərindən | **`şəyirdlər-in-dən`** | 6.0 | `şəyirdlər` |
475
- | yeməklərindən | **`yeməklər-in-dən`** | 6.0 | `yeməklər` |
476
- | qalxandan | **`qalx-an-dan`** | 6.0 | `qalx` |
477
- | kitabından | **`kitab-ın-dan`** | 6.0 | `kitab` |
478
- | açarından | **`açar-ın-dan`** | 6.0 | `açar` |
479
- | patriarxından | **`patriarx-ın-dan`** | 6.0 | `patriarx` |
480
- | gətirəndən | **`gətir-ən-dən`** | 6.0 | `gətir` |
481
- | qadağadan | **`qadağa-dan`** | 4.5 | `qadağa` |
482
- | ştirlisin | **`ştirlis-in`** | 4.5 | `ştirlis` |
483
- | fikirlərinin | **`fikirləri-nin`** | 4.5 | `fikirləri` |
484
- | təyyarəsinin | **`təyyarəsi-nin`** | 4.5 | `təyyarəsi` |
485
- | frenkinin | **`frenki-nin`** | 4.5 | `frenki` |
486
- | məzmundan | **`məzmun-dan`** | 4.5 | `məzmun` |
487
- | aerodinamikanın | **`aerodinamika-nın`** | 4.5 | `aerodinamika` |
488
- | intonasiyalardan | **`intonasiyalar-dan`** | 4.5 | `intonasiyalar` |
489
 
490
  ### 6.6 Linguistic Interpretation
491
 
492
  > **Automated Insight:**
493
- The language AZ 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.
494
 
495
  ---
496
  ## 7. Summary & Recommendations
@@ -502,7 +537,7 @@ The language AZ appears to be more isolating or has a highly fixed vocabulary. W
502
  | Component | Recommended | Rationale |
503
  |-----------|-------------|-----------|
504
  | Tokenizer | **64k BPE** | Best compression (5.13x) |
505
- | N-gram | **2-gram** | Lowest perplexity (405) |
506
  | Markov | **Context-4** | Highest predictability (96.5%) |
507
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
508
 
@@ -717,4 +752,4 @@ MIT License - Free for academic and commercial use.
717
  ---
718
  *Generated by Wikilangs Models Pipeline*
719
 
720
- *Report Date: 2026-01-03 09:50:56*
 
1
  ---
2
  language: az
3
+ language_name: Azerbaijani
4
  language_family: turkic_oghuz
5
  tags:
6
  - wikilangs
 
10
  - n-gram
11
  - markov
12
  - wikipedia
13
+ - feature-extraction
14
+ - sentence-similarity
15
+ - tokenization
16
+ - n-grams
17
+ - markov-chain
18
+ - text-mining
19
+ - fasttext
20
+ - babelvec
21
+ - vocabulous
22
+ - vocabulary
23
  - monolingual
24
  - family-turkic_oghuz
25
  license: mit
26
  library_name: wikilangs
27
+ pipeline_tag: text-generation
28
  datasets:
29
  - omarkamali/wikipedia-monthly
30
  dataset_info:
 
33
  metrics:
34
  - name: best_compression_ratio
35
  type: compression
36
+ value: 5.131
37
  - name: best_isotropy
38
  type: isotropy
39
+ value: 0.8140
40
  - name: vocabulary_size
41
  type: vocab
42
  value: 0
43
+ generated: 2026-01-04
44
  ---
45
 
46
+ # Azerbaijani - Wikilangs Models
47
  ## Comprehensive Research Report & Full Ablation Study
48
 
49
+ This repository contains NLP models trained and evaluated by Wikilangs, specifically on **Azerbaijani** Wikipedia data.
50
  We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and word embeddings.
51
 
52
  ## 📋 Repository Contents
 
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)
 
90
 
91
  | Vocab Size | Compression | Avg Token Len | UNK Rate | Total Tokens |
92
  |------------|-------------|---------------|----------|--------------|
93
+ | **8k** | 3.945x | 3.95 | 0.0962% | 1,248,644 |
94
+ | **16k** | 4.426x | 4.43 | 0.1079% | 1,113,127 |
95
+ | **32k** | 4.825x | 4.83 | 0.1176% | 1,021,125 |
96
+ | **64k** | 5.131x 🏆 | 5.13 | 0.1251% | 960,074 |
97
 
98
  ### Tokenization Examples
99
 
100
  Below are sample sentences tokenized with each vocabulary size:
101
 
102
+ **Sample 1:** `()aləminin dəstəsinin fəsiləsinə aid bitki cinsi. Sinonimləri Heterotipik sin...`
 
 
 
 
 
 
 
 
 
103
 
104
  | Vocab | Tokens | Count |
105
  |-------|--------|-------|
 
108
  | 32k | `▁() ▁— ▁aləminin ▁dəstəsinin ▁fəsiləsinə ▁aid ▁bitki ▁cinsi . ▁sinonimləri ... (+6 more)` | 16 |
109
  | 64k | `▁() ▁— ▁aləminin ▁dəstəsinin ▁fəsiləsinə ▁aid ▁bitki ▁cinsi . ▁sinonimləri ... (+6 more)` | 16 |
110
 
111
+ **Sample 2:** `() aləminin dəstəsinin fəsiləsinin cinsinə aid bitki növü. Sinonimləri Homotip...`
112
 
113
  | Vocab | Tokens | Count |
114
  |-------|--------|-------|
115
+ | 8k | `▁() ▁— aləminin ▁dəstəsininfəsiləsinin ▁cinsinə ▁aid ▁bitkinövü . ... (+8 more)` | 18 |
116
+ | 16k | `▁() ▁— aləminin ▁dəstəsininfəsiləsinin ▁cinsinə ▁aid ▁bitkinövü . ... (+8 more)` | 18 |
117
+ | 32k | `▁() ▁— aləminindəstəsinin ▁fəsiləsinin ▁cinsinə ▁aidbitkinövü . ... (+8 more)` | 18 |
118
+ | 64k | `▁() ▁— aləminindəstəsinin ▁fəsiləsinin ▁cinsinə ▁aidbitkinövü . ... (+8 more)` | 18 |
119
+
120
+ **Sample 3:** `.lr — Liberiyanın internet kodu. Xarici keçidlər IANA .lr whois information səvi...`
121
+
122
+ | Vocab | Tokens | Count |
123
+ |-------|--------|-------|
124
+ | 8k | `▁. l r ▁— ▁li ber iyanın ▁internet ▁kodu . ... (+18 more)` | 28 |
125
+ | 16k | `▁. l r ▁— ▁liber iyanın ▁internet ▁kodu . ▁xarici ... (+13 more)` | 23 |
126
+ | 32k | `▁. lr ▁— ▁liber iyanın ▁internet ▁kodu . ▁xarici ▁keçidlər ... (+8 more)` | 18 |
127
+ | 64k | `▁. lr ▁— ▁liber iyanın ▁internet ▁kodu . ▁xarici ▁keçidlər ... (+8 more)` | 18 |
128
 
129
 
130
  ### Key Findings
131
 
132
+ - **Best Compression:** 64k achieves 5.131x compression
133
+ - **Lowest UNK Rate:** 8k with 0.0962% unknown tokens
134
  - **Trade-off:** Larger vocabularies improve compression but increase model size
135
  - **Recommendation:** 32k vocabulary provides optimal balance for production use
136
 
 
147
 
148
  | N-gram | Variant | Perplexity | Entropy | Unique N-grams | Top-100 Coverage | Top-1000 Coverage |
149
  |--------|---------|------------|---------|----------------|------------------|-------------------|
150
+ | **2-gram** | Word | 267,397 | 18.03 | 1,224,963 | 4.8% | 13.7% |
151
+ | **2-gram** | Subword | 404 🏆 | 8.66 | 18,219 | 58.1% | 97.7% |
152
+ | **3-gram** | Word | 584,031 | 19.16 | 1,748,154 | 4.1% | 9.8% |
153
+ | **3-gram** | Subword | 3,741 | 11.87 | 158,841 | 20.7% | 61.1% |
154
+ | **4-gram** | Word | 1,231,291 | 20.23 | 3,034,353 | 3.9% | 8.4% |
155
+ | **4-gram** | Subword | 21,126 | 14.37 | 962,195 | 10.3% | 32.7% |
156
+ | **5-gram** | Word | 931,111 | 19.83 | 2,270,890 | 4.5% | 9.8% |
157
+ | **5-gram** | Subword | 81,852 | 16.32 | 3,259,009 | 6.2% | 20.7% |
158
 
159
  ### Top 5 N-grams by Size
160
 
 
162
 
163
  | Rank | N-gram | Count |
164
  |------|--------|-------|
165
+ | 1 | `və ya` | 84,279 |
166
+ | 2 | `xarici keçidlər` | 65,570 |
167
+ | 3 | `həmçinin bax` | 61,824 |
168
+ | 4 | `i̇stinadlar xarici` | 45,903 |
169
+ | 5 | `i̇stinadlar həmçinin` | 30,953 |
170
 
171
  **3-grams (Word):**
172
 
173
  | Rank | N-gram | Count |
174
  |------|--------|-------|
175
+ | 1 | `i̇stinadlar xarici keçidlər` | 45,411 |
176
+ | 2 | `i̇stinadlar həmçinin bax` | 30,925 |
177
+ | 3 | `fəsiləsinin cinsinə aid` | 20,614 |
178
+ | 4 | `dəstəsinin fəsiləsinin cinsinə` | 18,390 |
179
+ | 5 | `aid bitki növü` | 17,244 |
180
 
181
  **4-grams (Word):**
182
 
183
  | Rank | N-gram | Count |
184
  |------|--------|-------|
185
+ | 1 | `dəstəsinin fəsiləsinin cinsinə aid` | 18,390 |
186
+ | 2 | `cinsinə aid bitki növü` | 17,225 |
187
+ | 3 | `fəsiləsinin cinsinə aid bitki` | 17,194 |
188
+ | 4 | `aləminin dəstəsinin fəsiləsinin cinsinə` | 14,711 |
189
+ | 5 | `növü i̇stinadlar həmçinin bax` | 10,186 |
190
+
191
+ **5-grams (Word):**
192
+
193
+ | Rank | N-gram | Count |
194
+ |------|--------|-------|
195
+ | 1 | `fəsiləsinin cinsinə aid bitki növü` | 17,191 |
196
+ | 2 | `dəstəsinin fəsiləsinin cinsinə aid bitki` | 15,001 |
197
+ | 3 | `aləminin dəstəsinin fəsiləsinin cinsinə aid` | 14,711 |
198
+ | 4 | `cinsinə aid bitki növü i̇stinadlar` | 9,355 |
199
+ | 5 | `yeni ümumi kataloqda qeydə alınmış` | 8,316 |
200
 
201
  **2-grams (Subword):**
202
 
203
  | Rank | N-gram | Count |
204
  |------|--------|-------|
205
+ | 1 | `n _` | 8,039,357 |
206
+ | 2 | `ə _` | 6,502,225 |
207
+ | 3 | `i n` | 6,211,070 |
208
+ | 4 | `a r` | 5,368,955 |
209
+ | 5 | `ə r` | 5,307,819 |
210
 
211
  **3-grams (Subword):**
212
 
213
  | Rank | N-gram | Count |
214
  |------|--------|-------|
215
+ | 1 | `l ə r` | 2,430,392 |
216
+ | 2 | `l a r` | 2,275,096 |
217
+ | 3 | `d ə _` | 2,158,334 |
218
+ | 4 | `i n _` | 2,041,519 |
219
+ | 5 | `a n _` | 1,830,488 |
220
 
221
  **4-grams (Subword):**
222
 
223
  | Rank | N-gram | Count |
224
  |------|--------|-------|
225
+ | 1 | `_ v ə _` | 1,480,720 |
226
+ | 2 | `l ə r i` | 1,249,750 |
227
+ | 3 | `l a r ı` | 1,061,145 |
228
+ | 4 | `i n d ə` | 1,055,926 |
229
+ | 5 | `n i n _` | 957,274 |
230
+
231
+ **5-grams (Subword):**
232
+
233
+ | Rank | N-gram | Count |
234
+ |------|--------|-------|
235
+ | 1 | `i n i n _` | 790,811 |
236
+ | 2 | `l ə r i n` | 652,788 |
237
+ | 3 | `i n d ə _` | 641,243 |
238
+ | 4 | `l a r ı n` | 574,577 |
239
+ | 5 | `ı n d a _` | 522,632 |
240
 
241
 
242
  ### Key Findings
243
 
244
+ - **Best Perplexity:** 2-gram (subword) with 404
245
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
246
+ - **Coverage:** Top-1000 patterns cover ~21% of corpus
247
  - **Recommendation:** 4-gram or 5-gram for best predictive performance
248
 
249
  ---
 
259
 
260
  | Context | Variant | Avg Entropy | Perplexity | Branching Factor | Unique Contexts | Predictability |
261
  |---------|---------|-------------|------------|------------------|-----------------|----------------|
262
+ | **1** | Word | 0.9399 | 1.918 | 11.42 | 1,720,154 | 6.0% |
263
+ | **1** | Subword | 1.1732 | 2.255 | 8.01 | 8,102 | 0.0% |
264
+ | **2** | Word | 0.3192 | 1.248 | 1.95 | 19,621,953 | 68.1% |
265
+ | **2** | Subword | 0.7463 | 1.678 | 5.27 | 64,909 | 25.4% |
266
+ | **3** | Word | 0.1046 | 1.075 | 1.20 | 38,212,993 | 89.5% |
267
+ | **3** | Subword | 0.8107 | 1.754 | 4.76 | 342,087 | 18.9% |
268
+ | **4** | Word | 0.0352 🏆 | 1.025 | 1.06 | 45,793,057 | 96.5% |
269
+ | **4** | Subword | 0.7288 | 1.657 | 3.64 | 1,627,867 | 27.1% |
270
 
271
  ### Generated Text Samples (Word-based)
272
 
 
274
 
275
  **Context Size 1:**
276
 
277
+ 1. `və 25 cilddə v əsr kilsələri keçmiş rodeziya adlı ilk britaniya həyat və proqramlar efir`
278
+ 2. `ildə fiziki cəhətdən əlverişsiz şərait yaratdı o təbriz universitetində asiya ölkələrinə marşal çini...`
279
+ 3. `ilə yenidən tamaşaya qoyur şirvanşahlar taxtında gözü ilə habelə qafqazın qərbi avropada genos...`
280
 
281
  **Context Size 2:**
282
 
283
+ 1. `və ya yalan olan bir cismin səthinin digər cismin səthi arasındakı əlaqəni araşdırır i̇sbat nəzəriyy...`
284
+ 2. `xarici keçidlər ssr xalq hərbi dəniz nazirinin köməkçisi içləyib ilin iyun ayında çimkent şəhəri res...`
285
+ 3. `i̇stinadlar xarici keçidlər yanvar kaltenbrunner bir parade videosu nuremberg duruşmasında kaltenbru...`
286
 
287
  **Context Size 3:**
288
 
289
+ 1. `i̇stinadlar xarici keçidlər profile at sport resutls org kişi velosipedçilər sürücüləri yay olimpiya...`
290
+ 2. `fəsiləsinin cinsinə aid bitki növü sinonimləri heterotipik sinonimləri i̇stinadlar həmçinin bax i̇ra...`
291
+ 3. `dəstəsinin fəsiləsinin cinsinə aid bitki növü i̇stinadlar həmçinin bax nizami süleymanov kərrar əbil...`
292
 
293
  **Context Size 4:**
294
 
295
+ 1. `dəstəsinin fəsiləsinin cinsinə aid heyvan növü i̇stinadlar həmçinin bax ildə təsvir edilən sərtqanad...`
296
+ 2. `cinsinə aid bitki növü təbii yayılması botaniki təsviri ekologiyası azərbaycanda yayılması i̇stifadə...`
297
+ 3. `fəsiləsinin cinsinə aid bitki növü i̇stinadlar həmçinin bax ildə təsvir edilən bitkilər ildə təsvir ...`
298
 
299
 
300
  ### Generated Text Samples (Subword-based)
 
303
 
304
  **Context Size 1:**
305
 
306
+ 1. `_sindırə,_ke_enı`
307
+ 2. `ak,_xşdinrmisə_i`
308
+ 3. `inı_bondəkilayaq`
309
 
310
  **Context Size 2:**
311
 
312
+ 1. `n_bələ_hüsymətliq`
313
+ 2. `ə_onlan_ehrə_il_m`
314
+ 3. `indlaşı_atınd_eds`
315
 
316
  **Context Size 3:**
317
 
318
+ 1. `lər_kuboku_olanmas`
319
+ 2. `lar._söz_əlaqədi_b`
320
+ 3. `də_yabr_ilə_yer,_r`
321
 
322
  **Context Size 4:**
323
 
324
+ 1. `_və_təhsili_ilə_çıx`
325
+ 2. `lərini_100_mində_il`
326
+ 3. `indən_yazdı,_lakin_`
327
 
328
 
329
  ### Key Findings
330
 
331
  - **Best Predictability:** Context-4 (word) with 96.5% predictability
332
  - **Branching Factor:** Decreases with context size (more deterministic)
333
+ - **Memory Trade-off:** Larger contexts require more storage (1,627,867 contexts)
334
  - **Recommendation:** Context-3 or Context-4 for text generation
335
 
336
  ---
 
346
 
347
  | Metric | Value |
348
  |--------|-------|
349
+ | Vocabulary Size | 756,239 |
350
+ | Total Tokens | 53,635,250 |
351
+ | Mean Frequency | 70.92 |
352
  | Median Frequency | 4 |
353
+ | Frequency Std Dev | 2293.39 |
354
 
355
  ### Most Common Words
356
 
357
  | Rank | Word | Frequency |
358
  |------|------|-----------|
359
+ | 1 | və | 1,485,732 |
360
+ | 2 | ildə | 413,531 |
361
+ | 3 | ilə | 412,011 |
362
+ | 4 | bir | 365,123 |
363
+ | 5 | bu | 360,987 |
364
+ | 6 | də | 230,701 |
365
+ | 7 | üçün | 222,167 |
366
+ | 8 | azərbaycan | 221,202 |
367
+ | 9 | olan | 220,810 |
368
+ | 10 | sonra | 181,029 |
369
 
370
  ### Least Common Words (from vocabulary)
371
 
372
  | Rank | Word | Frequency |
373
  |------|------|-----------|
374
+ | 1 | gallaghers | 2 |
375
+ | 2 | liamın | 2 |
376
+ | 3 | liamla | 2 |
377
+ | 4 | backstab | 2 |
378
+ | 5 | antonioi | 2 |
379
+ | 6 | nipissinq | 2 |
380
+ | 7 | votivkirche | 2 |
381
+ | 8 | pirtle | 2 |
382
+ | 9 | takaxasinin | 2 |
383
+ | 10 | caporael | 2 |
384
 
385
  ### Zipf's Law Analysis
386
 
387
  | Metric | Value |
388
  |--------|-------|
389
  | Zipf Coefficient | 0.9645 |
390
+ | R² (Goodness of Fit) | 0.992387 |
391
  | Adherence Quality | **excellent** |
392
 
393
  ### Coverage Analysis
394
 
395
  | Top N Words | Coverage |
396
  |-------------|----------|
397
+ | Top 100 | 20.8% |
398
  | Top 1,000 | 45.3% |
399
+ | Top 5,000 | 65.5% |
400
  | Top 10,000 | 73.7% |
401
 
402
  ### Key Findings
403
 
404
+ - **Zipf Compliance:** R²=0.9924 indicates excellent adherence to Zipf's law
405
+ - **High Frequency Dominance:** Top 100 words cover 20.8% of corpus
406
+ - **Long Tail:** 746,239 words needed for remaining 26.3% coverage
407
 
408
  ---
409
  ## 5. Word Embeddings Evaluation
 
419
 
420
  ### 5.1 Cross-Lingual Alignment
421
 
422
+ ![Alignment Quality](visualizations/embedding_alignment_quality.png)
423
+
424
+ ![Multilingual t-SNE](visualizations/embedding_tsne_multilingual.png)
425
 
426
 
427
  ### 5.2 Model Comparison
428
 
429
  | Model | Dimension | Isotropy | Semantic Density | Alignment R@1 | Alignment R@10 |
430
  |-------|-----------|----------|------------------|---------------|----------------|
431
+ | **mono_32d** | 32 | 0.8140 🏆 | 0.3681 | N/A | N/A |
432
+ | **mono_64d** | 64 | 0.8077 | 0.2833 | N/A | N/A |
433
+ | **mono_128d** | 128 | 0.7661 | 0.2223 | N/A | N/A |
434
+ | **aligned_32d** | 32 | 0.8140 | 0.3594 | 0.1680 | 0.4820 |
435
+ | **aligned_64d** | 64 | 0.8077 | 0.2928 | 0.2820 | 0.7100 |
436
+ | **aligned_128d** | 128 | 0.7661 | 0.2246 | 0.4440 | 0.7780 |
437
 
438
  ### Key Findings
439
 
440
+ - **Best Isotropy:** mono_32d with 0.8140 (more uniform distribution)
441
+ - **Semantic Density:** Average pairwise similarity of 0.2918. Lower values indicate better semantic separation.
442
+ - **Alignment Quality:** Aligned models achieve up to 44.4% R@1 in cross-lingual retrieval.
443
  - **Recommendation:** 128d aligned for best cross-lingual performance
444
 
445
  ---
446
  ## 6. Morphological Analysis (Experimental)
447
 
 
 
448
  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.
449
 
450
  ### 6.1 Productivity & Complexity
451
 
452
  | Metric | Value | Interpretation | Recommendation |
453
  |--------|-------|----------------|----------------|
454
+ | Productivity Index | **5.000** | High morphological productivity | Reliable analysis |
455
+ | Idiomaticity Gap | **-0.527** | Low formulaic content | - |
456
 
457
  ### 6.2 Affix Inventory (Productive Units)
458
 
 
465
  #### Productive Suffixes
466
  | Suffix | Examples |
467
  |--------|----------|
468
+ | `-n` | kinopovestin, kristofferson, morfologiyasının |
469
+ | `-a` | metraja, irradiyasiya, razumovskaya |
470
+ | `-in` | kinopovestin, kriolitin, şikin |
471
+ | `-ın` | morfologiyasının, başın, buxtaların |
472
+ | `-an` | mozaikasından, qaçmazdan, tsiklopropan |
473
+ | `-ar` | vəzifəsimajoritar, yaratmışlar, tubalar |
474
+ | `-ən` | pərakəndəliyindən, gərginləşməsindən, birincidən |
475
+ | `-nın` | morfologiyasının, tistanın, andrianın |
476
 
477
  ### 6.3 Bound Stems (Lexical Roots)
478
 
 
480
 
481
  | Stem | Cohesion | Substitutability | Examples |
482
  |------|----------|------------------|----------|
483
+ | `ərba` | 2.70x | 42 contexts | ərbaa, ərbab, lərba |
484
+ | `rbay` | 2.38x | 53 contexts | orbay, arbay, erbay |
485
+ | `arix` | 2.17x | 73 contexts | larix, tarix, farix |
486
+ | `ayca` | 2.82x | 24 contexts | cayca, tayca, sayca |
487
+ | `mişd` | 1.65x | 164 contexts | mişdi, emişdi, mişdir |
488
+ | `nlar` | 1.37x | 429 contexts | anlar, nları, onlar |
489
+ | `ərəf` | 1.80x | 86 contexts | şərəf, ərəfə, tərəf |
490
+ | `lmiş` | 1.76x | 94 contexts | ölmiş, almiş, olmiş |
491
+ | `mışd` | 1.60x | 142 contexts | mışdı, mışdır, camışda |
492
+ | `ycan` | 2.94x | 13 contexts | aycan, bəycan, beycan |
493
+ | `qlar` | 1.45x | 196 contexts | aqlar, qlarn, lıqlar |
494
+ | `əfin` | 1.66x | 97 contexts | rəfin, dəfin, səfinə |
495
 
496
  ### 6.4 Affix Compatibility (Co-occurrence)
497
 
 
506
 
507
  | Word | Suggested Split | Confidence | Stem |
508
  |------|-----------------|------------|------|
509
+ | foneminin | **`fonem-in-in`** | 6.0 | `fonem` |
510
+ | təmsillərinin | **`təmsillər-in-in`** | 6.0 | `təmsillər` |
511
+ | qətiyyətinin | **`qətiyyət-in-in`** | 6.0 | `qətiyyət` |
512
+ | büküşlərinin | **`büküşlər-in-in`** | 6.0 | `büküşlər` |
513
+ | hədisçilərinin | **`hədisçilər-in-in`** | 6.0 | `hədisçilər` |
514
+ | planlaşdırmaqda | **`planlaşdırmaq-da`** | 4.5 | `planlaşdırmaq` |
515
+ | bölmələrimizin | **`bölmələrimiz-in`** | 4.5 | `bölmələrimiz` |
516
+ | heteranın | **`hetera-nın`** | 4.5 | `hetera` |
517
+ | somervillin | **`somervill-in`** | 4.5 | `somervill` |
518
+ | tanımanın | **`tanıma-nın`** | 4.5 | `tanıma` |
519
+ | meyitlərin | **`meyitlər-in`** | 4.5 | `meyitlər` |
520
+ | kameralizmin | **`kameralizm-in`** | 4.5 | `kameralizm` |
521
+ | burnettin | **`burnett-in`** | 4.5 | `burnett` |
522
+ | mussadıqın | **`mussadıq-ın`** | 4.5 | `mussadıq` |
523
+ | qalaçanın | **`qalaça-nın`** | 4.5 | `qalaça` |
524
 
525
  ### 6.6 Linguistic Interpretation
526
 
527
  > **Automated Insight:**
528
+ The language Azerbaijani shows high morphological productivity. The subword models are significantly more efficient than word models, suggesting a rich system of affixation or compounding.
529
 
530
  ---
531
  ## 7. Summary & Recommendations
 
537
  | Component | Recommended | Rationale |
538
  |-----------|-------------|-----------|
539
  | Tokenizer | **64k BPE** | Best compression (5.13x) |
540
+ | N-gram | **2-gram** | Lowest perplexity (404) |
541
  | Markov | **Context-4** | Highest predictability (96.5%) |
542
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
543
 
 
752
  ---
753
  *Generated by Wikilangs Models Pipeline*
754
 
755
+ *Report Date: 2026-01-04 14:36:36*
models/embeddings/aligned/az_128d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b46ba1b052b35932bc6f51e2fe778de34960a65c3baa640f6ed25c40dab3b5d3
3
+ size 1527436825
models/embeddings/aligned/az_128d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "az", "dim": 128, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/az_128d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2439fe0a1148bf99a41df0c37a0ce3171d6bf61a8eb45bd372461116a47ac7d4
3
+ size 65664
models/embeddings/aligned/az_128d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "az",
3
+ "dimension": 128,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 84414,
7
+ "vocab_size": 482300
8
+ }
models/embeddings/aligned/az_32d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf83be887640b63db335f2ea2b265802460f71e1b5d35df94c51b678b62746d5
3
+ size 389030425
models/embeddings/aligned/az_32d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "az", "dim": 32, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/az_32d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:161b2536a9024040d0a0faad847463eb203576a5211e7dc8dad1f611de0c7ec3
3
+ size 4224
models/embeddings/aligned/az_32d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "az",
3
+ "dimension": 32,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 84414,
7
+ "vocab_size": 482300
8
+ }
models/embeddings/aligned/az_64d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:04403b5768b23d9cc4fbf7e48dddd21508a562afa20c0faed9a5206599d45d8b
3
+ size 768499225
models/embeddings/aligned/az_64d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "az", "dim": 64, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/az_64d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5c3092c75598cf4cfd0df755b1a8b05892d8f9a44d0bd52e5b4a52e669dca9cb
3
+ size 16512
models/embeddings/aligned/az_64d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "az",
3
+ "dimension": 64,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 84414,
7
+ "vocab_size": 482300
8
+ }
models/embeddings/monolingual/az_128d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5d1aa42ac742f348af9cadb50b55f34e366914efb75ea61aa97a9c20e03e13b6
3
- size 1525679150
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b46ba1b052b35932bc6f51e2fe778de34960a65c3baa640f6ed25c40dab3b5d3
3
+ size 1527436825
models/embeddings/monolingual/az_128d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 128
13
  },
14
- "vocab_size": 480615
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 128
13
  },
14
+ "vocab_size": 482300
15
  }
models/embeddings/monolingual/az_32d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d370a43819df505438eb10c3e53922a704d1a21f86aab04f73bbd33be8dd0075
3
- size 388566830
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf83be887640b63db335f2ea2b265802460f71e1b5d35df94c51b678b62746d5
3
+ size 389030425
models/embeddings/monolingual/az_32d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 32
13
  },
14
- "vocab_size": 480615
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 32
13
  },
14
+ "vocab_size": 482300
15
  }
models/embeddings/monolingual/az_64d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4072e85e7840ba265c42156dd1d4f2342da7fb62cb52ac21d1b6fdbb2ecc2fba
3
- size 767604270
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:04403b5768b23d9cc4fbf7e48dddd21508a562afa20c0faed9a5206599d45d8b
3
+ size 768499225
models/embeddings/monolingual/az_64d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 64
13
  },
14
- "vocab_size": 480615
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 64
13
  },
14
+ "vocab_size": 482300
15
  }
models/subword_markov/az_markov_ctx1_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6c7409ea67865357c325ca7f791fa272f8dad5c303f2dca201b90c8247483c09
3
- size 458216
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b79afae90d5b990021e1abbebe8f39e487ee4b1fb8a1fa2e42e020b052b083d
3
+ size 455906
models/subword_markov/az_markov_ctx1_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_contexts": 8084,
6
- "total_transitions": 408780568
7
  }
 
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_contexts": 8102,
6
+ "total_transitions": 411540446
7
  }
models/subword_markov/az_markov_ctx2_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4d5864cd1185e963b662764ec1b090edf953a69e9dc6e006716f24e0b3428d9c
3
- size 2804436
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ae491954efd141eb449d0607575f9a8d6ea3ebdddafc2a11bcf0c0f5afba28a
3
+ size 2765159
models/subword_markov/az_markov_ctx2_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_contexts": 64659,
6
- "total_transitions": 408574228
7
  }
 
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_contexts": 64909,
6
+ "total_transitions": 411332483
7
  }
models/subword_markov/az_markov_ctx3_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fbfe0023ffc34f51e9adb5b0123c302c4505ca0f78d9903a5353ad83bc040ca2
3
- size 13127533
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:543c4822ffc5c1152310fec0689fd7b4c2feff03ae031aeca69cc41b0b2dd182
3
+ size 13156257
models/subword_markov/az_markov_ctx3_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_contexts": 341910,
6
- "total_transitions": 408367888
7
  }
 
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_contexts": 342087,
6
+ "total_transitions": 411124520
7
  }
models/subword_markov/az_markov_ctx4_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d1d38b5ce80eb4dbc53717d081e6f5c463f70eb15ac6704164d701a01b5d73c8
3
- size 47177591
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c70e92c5dcd635adfdcd222a690ede60f9a274e27fe6e74fe3ee9158b8df54eb
3
+ size 47266573
models/subword_markov/az_markov_ctx4_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_contexts": 1630545,
6
- "total_transitions": 408161548
7
  }
 
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_contexts": 1627867,
6
+ "total_transitions": 410916557
7
  }
models/subword_ngram/az_2gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:01b117009956861092e54c0b117771ba989b4436c6acc79c1f645ff057025622
3
- size 254254
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e4094a2cfd09d96d736ee20e6dedab981a8b31598ff00fbb8f28137eb84f1e8
3
+ size 255378
models/subword_ngram/az_2gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_ngrams": 18177,
6
- "total_ngrams": 408780568
7
  }
 
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_ngrams": 18219,
6
+ "total_ngrams": 411540446
7
  }
models/subword_ngram/az_3gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4e107d66818a354fecbd66a36ed7ba07ddeb4d4ba1af9def9601f9c38123b5bf
3
- size 2033861
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:528cc5e08e946322542308a651296e68fda8e7f2c4d463ef2e8ce4120e601566
3
+ size 2016526
models/subword_ngram/az_3gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_ngrams": 159097,
6
- "total_ngrams": 408574228
7
  }
 
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_ngrams": 158841,
6
+ "total_ngrams": 411332483
7
  }
models/subword_ngram/az_4gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:77a17978a0469bc6812197f3796b161e5474248d30fdc332daa5a855081be2ab
3
- size 12006457
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1742dbb434165bff2827355e118af4471ae00e1c77f13ae31524b61932360f7
3
+ size 11969438
models/subword_ngram/az_4gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "az",
5
- "unique_ngrams": 964243,
6
- "total_ngrams": 408367888
7
  }
 
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "az",
5
+ "unique_ngrams": 962195,
6
+ "total_ngrams": 411124520
7
  }
models/subword_ngram/az_5gram_subword.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70529944a2ad79220a459b936cc804e2c2766fe75e8bc1745a542fb7c66fa523
3
+ size 41203229
models/subword_ngram/az_5gram_subword_metadata.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "n": 5,
3
+ "variant": "subword",
4
+ "language": "az",
5
+ "unique_ngrams": 3259009,
6
+ "total_ngrams": 410916557
7
+ }
models/tokenizer/az_tokenizer_16k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5d3bccd20c5d35def160ef9f75d6c037555118c80fb3b040300f3fde4b6134ad
3
- size 525178
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3430004f2f06a739f24683cede680b53cdb396520eef92087708aaed5b9c488c
3
+ size 525300
models/tokenizer/az_tokenizer_16k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/az_tokenizer_32k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fd39d1d92e7df1344fde819e809b74326c591787c873fa720a5fd8890b719f47
3
- size 827135
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e92388cb43f9f6024f3e63fcdc71ae4a4373847f06cc4b3591f17fc241a69358
3
+ size 827326
models/tokenizer/az_tokenizer_32k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/az_tokenizer_64k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a276acb692e24539e0cc305ad0e3f3f4e3933a990782bddba5a70cd50a37d340
3
- size 1442474
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7ee32dd13a6a85672bd32399004d02f25c58a359b51760790769f7e1862514b5
3
+ size 1443741
models/tokenizer/az_tokenizer_64k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/az_tokenizer_8k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c4cea7e32c79e0a8f2372c125bd6ffd15c6db94b9860203f1518d4710f23cd59
3
- size 379373
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:510cce948124ac5f2f6aa61bb41305eba0cd4c63cfe3bda43327caa6f06207d0
3
+ size 379370
models/tokenizer/az_tokenizer_8k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/vocabulary/az_vocabulary.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:c5a35d8cbf3baef9f4f72d17c5715a1606edf5cd9c35ab2b583aeb2454944099
3
- size 11798895
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa0ee519296464675a9036bd37b6e4883825e3411db8bec28340f6187f95b395
3
+ size 11839727
models/vocabulary/az_vocabulary_metadata.json CHANGED
@@ -1,17 +1,17 @@
1
  {
2
  "language": "az",
3
- "vocabulary_size": 753394,
4
  "variant": "full",
5
  "statistics": {
6
- "type_token_ratio": 0.03161422652176624,
7
  "coverage": {
8
- "top_100": 0.20346273052205324,
9
- "top_1000": 0.44472442805281387,
10
- "top_5000": 0.6427898397241532,
11
- "top_10000": 0.7236870151642821
12
  },
13
- "hapax_count": 961452,
14
- "hapax_ratio": 0.560663756395618,
15
- "total_documents": 206340
16
  }
17
  }
 
1
  {
2
  "language": "az",
3
+ "vocabulary_size": 756239,
4
  "variant": "full",
5
  "statistics": {
6
+ "type_token_ratio": 0.03151627333092637,
7
  "coverage": {
8
+ "top_100": 0.2039517988648785,
9
+ "top_1000": 0.44508742002007223,
10
+ "top_5000": 0.6429971263810469,
11
+ "top_10000": 0.7238551252382953
12
  },
13
+ "hapax_count": 964543,
14
+ "hapax_ratio": 0.5605259701693764,
15
+ "total_documents": 207963
16
  }
17
  }
models/word_markov/az_markov_ctx1_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:552ab4c9e20cc54d226359a00acea16bbe48ee1f46733923ca6db57110f05da8
3
- size 189452830
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:32f9e233daa168a3e48f58fdb17214ff8cdc9dde1b24f27b86692a416e439162
3
+ size 190193554
models/word_markov/az_markov_ctx1_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "az",
5
- "unique_contexts": 1714220,
6
- "total_transitions": 54036518
7
  }
 
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "az",
5
+ "unique_contexts": 1720154,
6
+ "total_transitions": 54391830
7
  }
models/word_markov/az_markov_ctx2_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d9e7acc98680210a34fe7dd9401dfa34d443ac51b5045bb300561df4c105c25c
3
- size 600844050
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:169557c3a748d25521981f5f6711a2dd553e89f3ba5927bf4dfeb5fae02404c6
3
+ size 603928174
models/word_markov/az_markov_ctx2_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "az",
5
- "unique_contexts": 19534498,
6
- "total_transitions": 53830178
7
  }
 
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "az",
5
+ "unique_contexts": 19621953,
6
+ "total_transitions": 54183867
7
  }