omarkamali commited on
Commit
6af90c4
·
verified ·
1 Parent(s): bb3819a

Upload all models and assets for azb (20251001)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. README.md +293 -133
  2. models/embeddings/monolingual/azb_128d.bin +2 -2
  3. models/embeddings/monolingual/azb_128d_metadata.json +5 -3
  4. models/embeddings/monolingual/azb_32d.bin +2 -2
  5. models/embeddings/monolingual/azb_32d_metadata.json +5 -3
  6. models/embeddings/monolingual/azb_64d.bin +2 -2
  7. models/embeddings/monolingual/azb_64d_metadata.json +5 -3
  8. models/subword_markov/azb_markov_ctx1_subword.parquet +2 -2
  9. models/subword_markov/azb_markov_ctx1_subword_metadata.json +2 -2
  10. models/subword_markov/azb_markov_ctx2_subword.parquet +2 -2
  11. models/subword_markov/azb_markov_ctx2_subword_metadata.json +2 -2
  12. models/subword_markov/azb_markov_ctx3_subword.parquet +2 -2
  13. models/subword_markov/azb_markov_ctx3_subword_metadata.json +2 -2
  14. models/subword_markov/azb_markov_ctx4_subword.parquet +2 -2
  15. models/subword_markov/azb_markov_ctx4_subword_metadata.json +2 -2
  16. models/subword_ngram/azb_2gram_subword.parquet +2 -2
  17. models/subword_ngram/azb_2gram_subword_metadata.json +2 -2
  18. models/subword_ngram/azb_3gram_subword.parquet +2 -2
  19. models/subword_ngram/azb_3gram_subword_metadata.json +2 -2
  20. models/subword_ngram/azb_4gram_subword.parquet +2 -2
  21. models/subword_ngram/azb_4gram_subword_metadata.json +2 -2
  22. models/tokenizer/azb_tokenizer_16k.model +2 -2
  23. models/tokenizer/azb_tokenizer_16k.vocab +0 -0
  24. models/tokenizer/azb_tokenizer_32k.model +2 -2
  25. models/tokenizer/azb_tokenizer_32k.vocab +0 -0
  26. models/tokenizer/azb_tokenizer_64k.model +2 -2
  27. models/tokenizer/azb_tokenizer_64k.vocab +0 -0
  28. models/tokenizer/azb_tokenizer_8k.model +2 -2
  29. models/tokenizer/azb_tokenizer_8k.vocab +0 -0
  30. models/vocabulary/azb_vocabulary.parquet +2 -2
  31. models/vocabulary/azb_vocabulary_metadata.json +10 -9
  32. models/word_markov/azb_markov_ctx1_word.parquet +2 -2
  33. models/word_markov/azb_markov_ctx1_word_metadata.json +2 -2
  34. models/word_markov/azb_markov_ctx2_word.parquet +2 -2
  35. models/word_markov/azb_markov_ctx2_word_metadata.json +2 -2
  36. models/word_markov/azb_markov_ctx3_word.parquet +2 -2
  37. models/word_markov/azb_markov_ctx3_word_metadata.json +2 -2
  38. models/word_markov/azb_markov_ctx4_word.parquet +2 -2
  39. models/word_markov/azb_markov_ctx4_word_metadata.json +2 -2
  40. models/word_ngram/azb_2gram_word.parquet +2 -2
  41. models/word_ngram/azb_2gram_word_metadata.json +2 -2
  42. models/word_ngram/azb_3gram_word.parquet +2 -2
  43. models/word_ngram/azb_3gram_word_metadata.json +2 -2
  44. models/word_ngram/azb_4gram_word.parquet +2 -2
  45. models/word_ngram/azb_4gram_word_metadata.json +2 -2
  46. visualizations/embedding_isotropy.png +0 -0
  47. visualizations/embedding_norms.png +0 -0
  48. visualizations/embedding_similarity.png +2 -2
  49. visualizations/markov_branching.png +0 -0
  50. visualizations/markov_contexts.png +0 -0
README.md CHANGED
@@ -23,14 +23,14 @@ dataset_info:
23
  metrics:
24
  - name: best_compression_ratio
25
  type: compression
26
- value: 4.198
27
  - name: best_isotropy
28
  type: isotropy
29
- value: 0.8242
30
  - name: vocabulary_size
31
  type: vocab
32
- value: 317640
33
- generated: 2025-12-27
34
  ---
35
 
36
  # AZB - Wikilangs Models
@@ -44,12 +44,13 @@ We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and
44
  ### Models & Assets
45
 
46
  - Tokenizers (8k, 16k, 32k, 64k)
47
- - N-gram models (2, 3, 4-gram)
48
- - Markov chains (context of 1, 2, 3 and 4)
49
  - Subword N-gram and Markov chains
50
- - Embeddings in various sizes and dimensions
51
  - Language Vocabulary
52
  - Language Statistics
 
53
  ![Performance Dashboard](visualizations/performance_dashboard.png)
54
 
55
  ### Analysis and Evaluation
@@ -59,7 +60,8 @@ We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and
59
  - [3. Markov Chain Evaluation](#3-markov-chain-evaluation)
60
  - [4. Vocabulary Analysis](#4-vocabulary-analysis)
61
  - [5. Word Embeddings Evaluation](#5-word-embeddings-evaluation)
62
- - [6. Summary & Recommendations](#6-summary--recommendations)
 
63
  - [Metrics Glossary](#appendix-metrics-glossary--interpretation-guide)
64
  - [Visualizations Index](#visualizations-index)
65
 
@@ -68,51 +70,57 @@ We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and
68
 
69
  ![Tokenizer Compression](visualizations/tokenizer_compression.png)
70
 
 
 
 
 
 
 
71
  ### Results
72
 
73
  | Vocab Size | Compression | Avg Token Len | UNK Rate | Total Tokens |
74
  |------------|-------------|---------------|----------|--------------|
75
- | **8k** | 3.179x | 3.12 | 0.3165% | 385,493 |
76
- | **16k** | 3.558x | 3.49 | 0.3542% | 344,465 |
77
- | **32k** | 3.895x | 3.82 | 0.3877% | 314,660 |
78
- | **64k** | 4.198x 🏆 | 4.12 | 0.4179% | 291,946 |
79
 
80
  ### Tokenization Examples
81
 
82
  Below are sample sentences tokenized with each vocabulary size:
83
 
84
- **Sample 1:** `ماقنتیتی ( ) روسیه اؤلکه‌سینده یئر آلان بیر کند دیر و مورمانسک اوبلاستیندا یئرلش...`
85
 
86
  | Vocab | Tokens | Count |
87
  |-------|--------|-------|
88
- | 8k | `▁م اق ن تی تی ▁( ▁) ▁روسیه ▁اؤلکه ▁سینده ... (+19 more)` | 29 |
89
- | 16k | `▁ماق ن تی تی ▁( ▁) ▁روسیه ▁اؤلکه ▁سینده ▁یئر ... (+16 more)` | 26 |
90
- | 32k | `▁ماق نتی تی ▁( ▁) ▁روسیه ▁اؤلکه ▁سینده ▁یئر ▁آلان ... (+13 more)` | 23 |
91
- | 64k | `▁ماق نتی تی ▁( ▁) ▁روسیه ▁اؤلکه ▁سینده ▁یئر ▁آلان ... (+13 more)` | 23 |
92
 
93
- **Sample 2:** `سید ضیاء الدین طباطبائی یزدی (دوغوم:۱۲۸۶ شیراز -اؤلوم:۷ شهریور ۱۳۴۸ تهران) احمدش...`
94
 
95
  | Vocab | Tokens | Count |
96
  |-------|--------|-------|
97
- | 8k | `▁سید ▁ض یا ء ▁الدین ▁ط با ط با ئی ... (+34 more)` | 44 |
98
- | 16k | `▁سید ▁ضیا ء ▁الدین ▁ط با ط با ئی ▁ی ... (+29 more)` | 39 |
99
- | 32k | `▁سید ▁ضیا ء ▁الدین ▁ط باط با ئی ▁یز دی ... (+26 more)` | 36 |
100
- | 64k | `▁سید ▁ضیاء ▁الدین ▁طباطبا ئی ▁یزدی ▁( دوغوم : ۱۲ ... (+21 more)` | 31 |
101
 
102
- **Sample 3:** `ویکسن (اینگیلیسجه: Vixen, Louisiana) آمریکانین لوئیزیانا ایالتینده یئرلشن بیر یا...`
103
 
104
  | Vocab | Tokens | Count |
105
  |-------|--------|-------|
106
- | 8k | `▁ویک سن ▁( اینگیلیسجه : v ix en , ▁louisiana ... (+15 more)` | 25 |
107
- | 16k | `▁ویک سن ▁( اینگیلیسجه : v ix en , ▁louisiana ... (+15 more)` | 25 |
108
- | 32k | `▁ویک سن ▁( اینگیلیسجه : v ix en , ▁louisiana ... (+15 more)` | 25 |
109
- | 64k | `▁ویک سن ▁( اینگیلیسجه : vixen , ▁louisiana ) ▁آمریکانین ... (+13 more)` | 23 |
110
 
111
 
112
  ### Key Findings
113
 
114
- - **Best Compression:** 64k achieves 4.198x compression
115
- - **Lowest UNK Rate:** 8k with 0.3165% unknown tokens
116
  - **Trade-off:** Larger vocabularies improve compression but increase model size
117
  - **Recommendation:** 32k vocabulary provides optimal balance for production use
118
 
@@ -121,57 +129,89 @@ Below are sample sentences tokenized with each vocabulary size:
121
 
122
  ![N-gram Perplexity](visualizations/ngram_perplexity.png)
123
 
 
 
124
  ![N-gram Coverage](visualizations/ngram_coverage.png)
125
 
126
  ### Results
127
 
128
- | N-gram | Perplexity | Entropy | Unique N-grams | Top-100 Coverage | Top-1000 Coverage |
129
- |--------|------------|---------|----------------|------------------|-------------------|
130
- | **2-gram** | 7,645 🏆 | 12.90 | 268,529 | 30.2% | 57.7% |
131
- | **2-gram** | 655 🏆 | 9.35 | 14,131 | 46.5% | 93.7% |
132
- | **3-gram** | 13,711 | 13.74 | 512,828 | 24.9% | 52.4% |
133
- | **3-gram** | 4,808 | 12.23 | 141,425 | 21.1% | 58.1% |
134
- | **4-gram** | 23,522 | 14.52 | 979,944 | 21.9% | 47.2% |
135
- | **4-gram** | 19,536 | 14.25 | 804,296 | 13.9% | 41.7% |
136
 
137
  ### Top 5 N-grams by Size
138
 
139
- **2-grams:**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
140
 
141
  | Rank | N-gram | Count |
142
  |------|--------|-------|
143
- | 1 | `بؤلمه :` | 391,479 |
144
- | 2 | `او ��` | 139,043 |
145
- | 3 | `‌ لار` | 137,076 |
146
- | 4 | `قایناقلار بؤلمه` | 125,059 |
147
- | 5 | `‌ نین` | 118,471 |
148
 
149
- **3-grams:**
150
 
151
  | Rank | N-gram | Count |
152
  |------|--------|-------|
153
- | 1 | `قایناقلار بؤلمه :` | 125,058 |
154
- | 2 | `( اینگیلیسجه :` | 112,489 |
155
- | 3 | `. قایناقلار بؤلمه` | 97,901 |
156
- | 4 | `قایناق لار` | 92,555 |
157
- | 5 | `ایشلدنلری طرفیندن یارانمیش` | 76,848 |
158
 
159
- **4-grams:**
160
 
161
  | Rank | N-gram | Count |
162
  |------|--------|-------|
163
- | 1 | `. قایناقلار بؤلمه :` | 97,901 |
164
- | 2 | `ایشلدنلری طرفیندن یارانمیش «` | 76,839 |
165
- | 3 | خلانیلیبدیر ) .` | 76,827 |
166
- | 4 | ، مقاله ‌` | 76,824 |
167
- | 5 | مقاله سیندن` | 76,824 |
168
 
169
 
170
  ### Key Findings
171
 
172
- - **Best Perplexity:** 2-gram with 655
173
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
174
- - **Coverage:** Top-1000 patterns cover ~42% of corpus
175
  - **Recommendation:** 4-gram or 5-gram for best predictive performance
176
 
177
  ---
@@ -179,55 +219,86 @@ Below are sample sentences tokenized with each vocabulary size:
179
 
180
  ![Markov Entropy](visualizations/markov_entropy.png)
181
 
 
 
182
  ![Markov Branching](visualizations/markov_branching.png)
183
 
184
  ### Results
185
 
186
- | Context | Avg Entropy | Perplexity | Branching Factor | Unique Contexts | Predictability |
187
- |---------|-------------|------------|------------------|-----------------|----------------|
188
- | **1** | 0.4591 | 1.375 | 4.28 | 1,005,885 | 54.1% |
189
- | **1** | 1.2059 | 2.307 | 10.74 | 2,934 | 0.0% |
190
- | **2** | 0.2308 | 1.173 | 1.71 | 4,302,287 | 76.9% |
191
- | **2** | 1.1061 | 2.153 | 8.32 | 31,498 | 0.0% |
192
- | **3** | 0.1101 | 1.079 | 1.28 | 7,369,809 | 89.0% |
193
- | **3** | 0.9393 | 1.918 | 5.09 | 262,151 | 6.1% |
194
- | **4** | 0.0637 🏆 | 1.045 | 1.14 | 9,387,722 | 93.6% |
195
- | **4** | 0.7222 🏆 | 1.650 | 3.21 | 1,333,406 | 27.8% |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
196
 
197
- ### Generated Text Samples
198
 
199
- Below are text samples generated from each Markov chain model:
 
 
200
 
201
  **Context Size 1:**
202
 
203
- 1. `‌ ایللیین اووللرین ده یاشایان اینسانلار بؤلمه : bhumibol adulyadej crown , footballer , alchimie du`
204
- 2. `. قان معائناتی و زنگین ‌ نین ایشلدنلری طرفیندن یارانمیش « panama 2 episodes ایلآدی رو`
205
- 3. `: sergey ryzhikov ( ۱۹۴۸ ایلینده گونئی کارولینا ( تحصیلات ) رومنی ( لهیستانجا : 2`
206
 
207
  **Context Size 2:**
208
 
209
- 1. `بؤلمه : وورت بؤلگه ‌ سیاراک بؤلگه ‌ سی ‌ دیر گؤرونتولر قایناق ‌ لار اینگیلیسجه ویکی`
210
- 2. `او ْ یونچو . ۲۶ ژوئن ۱۹۵۷ میلادی تاریخیندا وفات ائدیب . قایناق ‌ لار اینگیلیسجه ویکی`
211
- 3. `‌ لار بؤلمه : آمریکا شهرلری بؤلمه : گرینزبورو دوغوملولار بؤلمه : ۲۰۱۹ - جو ایلینده افغانیستاندا`
212
 
213
  **Context Size 3:**
214
 
215
- 1. `قایناقلار بؤلمه : هیندوستان کندلری en : kahra`
216
- 2. `( اینگیلیسجه : louis prima ) آمریکالی موغنی ، یازیچی و او ْ ندان ایستیفاده ائتمیشدیر . روبنسین`
217
- 3. `. قایناقلار بؤلمه : هیندوستان کندلری en : kalawad`
218
 
219
  **Context Size 4:**
220
 
221
- 1. `. قایناقلار بؤلمه : هیندوستان کندلری en : balowal`
222
- 2. `ایشلدنلری طرفیندن یارانمیش « crabtree ' s catalyst » ، مقاله ‌ سیندن گؤتورولوبدور . ( ۷ سپتامبر ۲۰۱۷`
223
- 3. `، مقاله ‌ سیندن گؤتورولوبدور . ( ۱۲ ژوئن ۲۰۱۸ تاریخینده یو ْ خلانیلیبدیر ) . بؤلمه : ایستادیوملار`
224
 
225
 
226
  ### Key Findings
227
 
228
- - **Best Predictability:** Context-4 with 93.6% predictability
229
  - **Branching Factor:** Decreases with context size (more deterministic)
230
- - **Memory Trade-off:** Larger contexts require more storage (1,333,406 contexts)
231
  - **Recommendation:** Context-3 or Context-4 for text generation
232
 
233
  ---
@@ -243,64 +314,64 @@ Below are text samples generated from each Markov chain model:
243
 
244
  | Metric | Value |
245
  |--------|-------|
246
- | Vocabulary Size | 317,640 |
247
- | Total Tokens | 16,697,106 |
248
- | Mean Frequency | 52.57 |
249
  | Median Frequency | 3 |
250
- | Frequency Std Dev | 1499.80 |
251
 
252
  ### Most Common Words
253
 
254
  | Rank | Word | Frequency |
255
  |------|------|-----------|
256
- | 1 | بؤلمه | 391,840 |
257
- | 2 | و | 291,285 |
258
- | 3 | اینگیلیسجه | 188,551 |
259
- | 4 | بیر | 169,796 |
260
- | 5 | او | 145,694 |
261
- | 6 | قایناقلار | 143,368 |
262
- | 7 | سی | 139,966 |
263
- | 8 | لار | 138,388 |
264
- | 9 | دیر | 132,038 |
265
- | 10 | the | 130,277 |
266
 
267
  ### Least Common Words (from vocabulary)
268
 
269
  | Rank | Word | Frequency |
270
  |------|------|-----------|
271
- | 1 | سئندن | 2 |
272
- | 2 | ائششکین | 2 |
273
- | 3 | onager | 2 |
274
- | 4 | داشناکلارلا | 2 |
275
- | 5 | قۇلان | 2 |
276
- | 6 | آسینۇس | 2 |
277
- | 7 | هئمیو | 2 |
278
- | 8 | تاپؽلمیشدیر | 2 |
279
- | 9 | kulan | 2 |
280
- | 10 | کسا | 2 |
281
 
282
  ### Zipf's Law Analysis
283
 
284
  | Metric | Value |
285
  |--------|-------|
286
- | Zipf Coefficient | 1.1842 |
287
- | R² (Goodness of Fit) | 0.995333 |
288
  | Adherence Quality | **excellent** |
289
 
290
  ### Coverage Analysis
291
 
292
  | Top N Words | Coverage |
293
  |-------------|----------|
294
- | Top 100 | 36.8% |
295
- | Top 1,000 | 66.7% |
296
- | Top 5,000 | 81.2% |
297
- | Top 10,000 | 85.9% |
298
 
299
  ### Key Findings
300
 
301
- - **Zipf Compliance:** R²=0.9953 indicates excellent adherence to Zipf's law
302
- - **High Frequency Dominance:** Top 100 words cover 36.8% of corpus
303
- - **Long Tail:** 307,640 words needed for remaining 14.1% coverage
304
 
305
  ---
306
  ## 5. Word Embeddings Evaluation
@@ -313,24 +384,110 @@ Below are text samples generated from each Markov chain model:
313
 
314
  ![t-SNE Sentences](visualizations/tsne_sentences.png)
315
 
316
- ### Model Comparison
317
 
318
- | Model | Vocab Size | Dimension | Avg Norm | Std Norm | Isotropy |
319
- |-------|------------|-----------|----------|----------|----------|
320
- | **mono_32d** | 134,712 | 32 | 5.098 | 1.243 | 0.8242 🏆 |
321
- | **mono_64d** | 134,712 | 64 | 5.525 | 1.165 | 0.7928 |
322
- | **mono_128d** | 134,712 | 128 | 6.056 | 1.052 | 0.7520 |
323
- | **embeddings_enhanced** | 0 | 0 | 0.000 | 0.000 | 0.0000 |
 
 
 
 
 
 
324
 
325
  ### Key Findings
326
 
327
- - **Best Isotropy:** mono_32d with 0.8242 (more uniform distribution)
328
- - **Dimension Trade-off:** Higher dimensions capture more semantics but reduce isotropy
329
- - **Vocabulary Coverage:** All models cover 134,712 words
330
- - **Recommendation:** 100d for balanced semantic capture and efficiency
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
331
 
332
  ---
333
- ## 6. Summary & Recommendations
334
 
335
  ![Performance Dashboard](visualizations/performance_dashboard.png)
336
 
@@ -338,11 +495,12 @@ Below are text samples generated from each Markov chain model:
338
 
339
  | Component | Recommended | Rationale |
340
  |-----------|-------------|-----------|
341
- | Tokenizer | **32k BPE** | Best compression (4.20x) with low UNK rate |
342
- | N-gram | **5-gram** | Lowest perplexity (655) |
343
- | Markov | **Context-4** | Highest predictability (93.6%) |
344
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
345
 
 
346
  ---
347
  ## Appendix: Metrics Glossary & Interpretation Guide
348
 
@@ -532,7 +690,8 @@ If you use these models in your research, please cite:
532
  author = {Kamali, Omar},
533
  title = {Wikilangs: Open NLP Models for Wikipedia Languages},
534
  year = {2025},
535
- publisher = {HuggingFace},
 
536
  url = {https://huggingface.co/wikilangs}
537
  institution = {Omneity Labs}
538
  }
@@ -548,7 +707,8 @@ MIT License - Free for academic and commercial use.
548
  - 🤗 Models: [huggingface.co/wikilangs](https://huggingface.co/wikilangs)
549
  - 📊 Data: [wikipedia-monthly](https://huggingface.co/datasets/omarkamali/wikipedia-monthly)
550
  - 👤 Author: [Omar Kamali](https://huggingface.co/omarkamali)
 
551
  ---
552
  *Generated by Wikilangs Models Pipeline*
553
 
554
- *Report Date: 2025-12-27 22:57:16*
 
23
  metrics:
24
  - name: best_compression_ratio
25
  type: compression
26
+ value: 4.148
27
  - name: best_isotropy
28
  type: isotropy
29
+ value: 0.8282
30
  - name: vocabulary_size
31
  type: vocab
32
+ value: 0
33
+ generated: 2026-01-03
34
  ---
35
 
36
  # AZB - Wikilangs Models
 
44
  ### Models & Assets
45
 
46
  - Tokenizers (8k, 16k, 32k, 64k)
47
+ - N-gram models (2, 3, 4, 5-gram)
48
+ - Markov chains (context of 1, 2, 3, 4 and 5)
49
  - Subword N-gram and Markov chains
50
+ - Embeddings in various sizes and dimensions (aligned and unaligned)
51
  - Language Vocabulary
52
  - Language Statistics
53
+
54
  ![Performance Dashboard](visualizations/performance_dashboard.png)
55
 
56
  ### Analysis and Evaluation
 
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)
67
 
 
70
 
71
  ![Tokenizer Compression](visualizations/tokenizer_compression.png)
72
 
73
+ ![Tokenizer Fertility](visualizations/tokenizer_fertility.png)
74
+
75
+ ![Tokenizer OOV](visualizations/tokenizer_oov.png)
76
+
77
+ ![Total Tokens](visualizations/tokenizer_total_tokens.png)
78
+
79
  ### Results
80
 
81
  | Vocab Size | Compression | Avg Token Len | UNK Rate | Total Tokens |
82
  |------------|-------------|---------------|----------|--------------|
83
+ | **8k** | 3.135x | 3.14 | 0.5011% | 364,218 |
84
+ | **16k** | 3.510x | 3.51 | 0.5610% | 325,334 |
85
+ | **32k** | 3.852x | 3.86 | 0.6157% | 296,427 |
86
+ | **64k** | 4.148x 🏆 | 4.15 | 0.6629% | 275,291 |
87
 
88
  ### Tokenization Examples
89
 
90
  Below are sample sentences tokenized with each vocabulary size:
91
 
92
+ **Sample 1:** `قالیکتیس (، ، ، ) ییٛرتیجیلار دسته‌سینه عایید حئیوان نؤعو. قایناقلار سیراسینا گؤ...`
93
 
94
  | Vocab | Tokens | Count |
95
  |-------|--------|-------|
96
+ | 8k | `▁قا لیک تیس ▁(، ▁، ▁، ▁) ▁ییٛرتیجیلار ▁دسته ▁سینه ... (+8 more)` | 18 |
97
+ | 16k | `▁قا لیک تیس ▁(، ▁، ▁، ▁) ▁ییٛرتیجیلار ▁دسته ▁سینه ... (+8 more)` | 18 |
98
+ | 32k | `▁قا لیک تیس ▁(، ▁، ▁، ▁) ▁ییٛرتیجیلار ▁دسته ▁سینه ... (+8 more)` | 18 |
99
+ | 64k | `▁قا لیک تیس ▁(، ▁، ▁، ▁) ▁ییٛرتیجیلار ▁دسته ▁سینه ... (+8 more)` | 18 |
100
 
101
+ **Sample 2:** `هیندوستان اؤلکه‌سینین کرالا ایالتینده بیر شهر دیر. بۇ شهرده مالایالم دیلی و اینگ...`
102
 
103
  | Vocab | Tokens | Count |
104
  |-------|--------|-------|
105
+ | 8k | `▁هیندوستان ▁اؤلکه ▁سینین ▁کرالا ▁ایالتینده ▁بیر ▁شهر ▁دیر . ▁بۇ ... (+10 more)` | 20 |
106
+ | 16k | `▁هیندوستان ▁اؤلکه ▁سینین ▁کرالا ▁ایالتینده ▁بیر ▁شهر ▁دیر . ▁بۇ ... (+10 more)` | 20 |
107
+ | 32k | `▁هیندوستان ▁اؤلکه ▁سینین ▁کرالا ▁ایالتینده ▁بیر ▁شهر ▁دیر . ▁بۇ ... (+10 more)` | 20 |
108
+ | 64k | `▁هیندوستان ▁اؤلکه ▁سینین ▁کرالا ▁ایالتینده ▁بیر ▁شهر ▁دیر . ▁بۇ ... (+10 more)` | 20 |
109
 
110
+ **Sample 3:** `آرقا, کارناتاکا Karnataka) هیندوستان اؤلکه‌سینین کارناتاکا ایالتینده بیر کند دیر...`
111
 
112
  | Vocab | Tokens | Count |
113
  |-------|--------|-------|
114
+ | 8k | `▁آر قا , ▁کارناتاکاkar n at aka ) ▁هیندوستان ... (+16 more)` | 26 |
115
+ | 16k | `▁آر قا , ▁کارناتاکاkar nat aka ) ▁هیندوستان ▁اؤلکه ... (+15 more)` | 25 |
116
+ | 32k | `▁آر قا , ▁کارناتاکاkarnataka ) ▁هیندوستان ▁اؤلکه ▁سینین ▁کارناتاکا ... (+13 more)` | 23 |
117
+ | 64k | `▁آر قا , ▁کارناتاکاkarnataka ) ▁هیندوستان ▁اؤلکه ▁سینین ▁کارناتاکا ... (+13 more)` | 23 |
118
 
119
 
120
  ### Key Findings
121
 
122
+ - **Best Compression:** 64k achieves 4.148x compression
123
+ - **Lowest UNK Rate:** 8k with 0.5011% unknown tokens
124
  - **Trade-off:** Larger vocabularies improve compression but increase model size
125
  - **Recommendation:** 32k vocabulary provides optimal balance for production use
126
 
 
129
 
130
  ![N-gram Perplexity](visualizations/ngram_perplexity.png)
131
 
132
+ ![N-gram Unique](visualizations/ngram_unique.png)
133
+
134
  ![N-gram Coverage](visualizations/ngram_coverage.png)
135
 
136
  ### Results
137
 
138
+ | N-gram | Variant | Perplexity | Entropy | Unique N-grams | Top-100 Coverage | Top-1000 Coverage |
139
+ |--------|---------|------------|---------|----------------|------------------|-------------------|
140
+ | **2-gram** | Word | 8,048 | 12.97 | 158,908 | 25.7% | 56.1% |
141
+ | **2-gram** | Subword | 528 🏆 | 9.04 | 12,648 | 51.6% | 95.7% |
142
+ | **3-gram** | Word | 10,249 | 13.32 | 236,749 | 22.6% | 53.6% |
143
+ | **3-gram** | Subword | 3,765 | 11.88 | 106,644 | 23.1% | 62.4% |
144
+ | **4-gram** | Word | 17,175 | 14.07 | 426,395 | 19.0% | 47.9% |
145
+ | **4-gram** | Subword | 15,100 | 13.88 | 581,225 | 14.6% | 44.8% |
146
 
147
  ### Top 5 N-grams by Size
148
 
149
+ **2-grams (Word):**
150
+
151
+ | Rank | N-gram | Count |
152
+ |------|--------|-------|
153
+ | 1 | `ایشلدنلری طرفیندن` | 75,584 |
154
+ | 2 | `مقاله‌سیندن گؤتورولوبدور` | 75,503 |
155
+ | 3 | `ویکی‌پدیاسی‌نین ایشلدنلری` | 73,734 |
156
+ | 4 | `اینگیلیسجه ویکی‌پدیاسی‌نین` | 71,132 |
157
+ | 5 | `قایناق‌لار اینگیلیسجه` | 70,880 |
158
+
159
+ **3-grams (Word):**
160
+
161
+ | Rank | N-gram | Count |
162
+ |------|--------|-------|
163
+ | 1 | `ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن` | 73,734 |
164
+ | 2 | `اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری` | 71,132 |
165
+ | 3 | `قایناق‌لار اینگیلیسجه ویکی‌پدیاسی‌نین` | 70,806 |
166
+ | 4 | `بیر یاشاییش منطقه‌سی‌دیر` | 40,399 |
167
+ | 5 | `بیر کند دیر` | 30,448 |
168
+
169
+ **4-grams (Word):**
170
+
171
+ | Rank | N-gram | Count |
172
+ |------|--------|-------|
173
+ | 1 | `اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن` | 71,132 |
174
+ | 2 | `قایناق‌لار اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری` | 70,806 |
175
+ | 3 | `سوْن نۆفوس ساییمی اساسيندا` | 24,568 |
176
+ | 4 | `شهرلرین لیستی قایناق‌لار اینگیلیسجه` | 22,937 |
177
+ | 5 | `لیستی قایناق‌لار اینگیلیسجه ویکی‌پدیاسی‌نین` | 22,937 |
178
+
179
+ **2-grams (Subword):**
180
 
181
  | Rank | N-gram | Count |
182
  |------|--------|-------|
183
+ | 1 | ن` | 1,868,991 |
184
+ | 2 | `_ ا` | 1,658,104 |
185
+ | 3 | _` | 1,437,263 |
186
+ | 4 | ی` | 1,393,221 |
187
+ | 5 | _` | 1,215,806 |
188
 
189
+ **3-grams (Subword):**
190
 
191
  | Rank | N-gram | Count |
192
  |------|--------|-------|
193
+ | 1 | `_ ا ی` | 717,380 |
194
+ | 2 | ن د` | 658,977 |
195
+ | 3 | ه _` | 585,522 |
196
+ | 4 | ا ر` | 580,226 |
197
+ | 5 | ی ن` | 470,621 |
198
 
199
+ **4-grams (Subword):**
200
 
201
  | Rank | N-gram | Count |
202
  |------|--------|-------|
203
+ | 1 | د ه _` | 347,347 |
204
+ | 2 | ا ر _` | 329,379 |
205
+ | 3 | ن د ه` | 320,994 |
206
+ | 4 | `_ ب ی ر` | 258,707 |
207
+ | 5 | ی ن _` | 257,628 |
208
 
209
 
210
  ### Key Findings
211
 
212
+ - **Best Perplexity:** 2-gram (subword) with 528
213
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
214
+ - **Coverage:** Top-1000 patterns cover ~45% of corpus
215
  - **Recommendation:** 4-gram or 5-gram for best predictive performance
216
 
217
  ---
 
219
 
220
  ![Markov Entropy](visualizations/markov_entropy.png)
221
 
222
+ ![Markov Contexts](visualizations/markov_contexts.png)
223
+
224
  ![Markov Branching](visualizations/markov_branching.png)
225
 
226
  ### Results
227
 
228
+ | Context | Variant | Avg Entropy | Perplexity | Branching Factor | Unique Contexts | Predictability |
229
+ |---------|---------|-------------|------------|------------------|-----------------|----------------|
230
+ | **1** | Word | 0.6640 | 1.584 | 5.09 | 726,930 | 33.6% |
231
+ | **1** | Subword | 1.0592 | 2.084 | 9.07 | 3,409 | 0.0% |
232
+ | **2** | Word | 0.1970 | 1.146 | 1.48 | 3,693,091 | 80.3% |
233
+ | **2** | Subword | 0.9308 | 1.906 | 6.57 | 30,905 | 6.9% |
234
+ | **3** | Word | 0.0689 | 1.049 | 1.14 | 5,447,170 | 93.1% |
235
+ | **3** | Subword | 0.8408 | 1.791 | 4.70 | 203,056 | 15.9% |
236
+ | **4** | Word | 0.0340 🏆 | 1.024 | 1.07 | 6,178,524 | 96.6% |
237
+ | **4** | Subword | 0.7000 | 1.625 | 3.22 | 953,883 | 30.0% |
238
+
239
+ ### Generated Text Samples (Word-based)
240
+
241
+ Below are text samples generated from each word-based Markov chain model:
242
+
243
+ **Context Size 1:**
244
+
245
+ 1. `و شمشددیل سلطانلیغی ایله برابر ایدی قایناقلار شهرلری آمریکا بیرلشمیش ایالتلرینین ایداری بؤلوملری ائش...`
246
+ 2. `بیر کند دیر و یا هچ اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن mountain a few years until 5`
247
+ 3. `اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن mountain wilbert minnesota مقاله‌سیندن گؤتورولوبدور ۸ آ...`
248
+
249
+ **Context Size 2:**
250
+
251
+ 1. `ایشلدنلری طرفیندن مقاله‌سیندن گؤتورولوبدور ۸ آقوست تاریخینده یوْخلانیلیبدیر شهرلری en güləh`
252
+ 2. `مقاله‌سیندن گؤتورولوبدور ۳۰ نوْوامبر تاریخینده یوْخلانیلیبدیر کولوبلاری en araks ararat fc مقاله‌سین...`
253
+ 3. `ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن indiana مقاله‌سیندن گؤتورولوبدور ۲۱ دسامبر تاریخینده یوْخلانیلیبدی...`
254
+
255
+ **Context Size 3:**
256
+
257
+ 1. `ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن مقاله‌سیندن گؤتورولوبدور ۳۰ نوْوامبر تاریخینده یوْخلانیلیبدیر گؤرو...`
258
+ 2. `اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن rutherfurd مقاله‌سیندن گؤتورولوبدور ۲۲ ژانویه تاریخینده...`
259
+ 3. `قایناق‌لار اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن مقاله‌سیندن گؤتورولوبدور ۱۹ جولای یوْخلانیلی...`
260
+
261
+ **Context Size 4:**
262
+
263
+ 1. `اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن مقاله‌سیندن گؤتورولوبدور ۸ آقوست تاریخینده یوْخلانیلیبد...`
264
+ 2. `قایناق‌لار اینگیلیسجه ویکی‌پدیاسی‌نین ایشلدنلری طرفیندن georgia مقاله‌سیندن گؤتورولوبدور ۸ آقوست تار...`
265
+ 3. `سوْن نۆفوس ساییمی اساسيندا ۱۳۷ نفر ایمیش و ویسوچینا اوستانیندا یئرلشیب بۆتون چک‌دا اوْلدوغو کیمی بۇ ...`
266
 
 
267
 
268
+ ### Generated Text Samples (Subword-based)
269
+
270
+ Below are text samples generated from each subword-based Markov chain model:
271
 
272
  **Context Size 1:**
273
 
274
+ 1. `_متیشاورله‌ستبشلی`
275
+ 2. `یتا_d_by_jottiep`
276
+ 3. `الاق‌لر)_ژادالين_`
277
 
278
  **Context Size 2:**
279
 
280
+ 1. `ین_حیازیل_دیکان_ا`
281
+ 2. `_ایلدا_۹_جی_giran`
282
+ 3. `ی_حافع_سان_آماری_`
283
 
284
  **Context Size 3:**
285
 
286
+ 1. `_ایدان_عوضوو._۲۷۳_`
287
+ 2. `ینده_روس_سال_منطقه‌`
288
+ 3. `ده_یوْخلانی_آما_خوب`
289
 
290
  **Context Size 4:**
291
 
292
+ 1. `نده_یوْخلانیلیبدیر).`
293
+ 2. `لار_یولو_۳۰_دسامبر_`
294
+ 3. `ینده_هر_گونئی_کاروا`
295
 
296
 
297
  ### Key Findings
298
 
299
+ - **Best Predictability:** Context-4 (word) with 96.6% predictability
300
  - **Branching Factor:** Decreases with context size (more deterministic)
301
+ - **Memory Trade-off:** Larger contexts require more storage (953,883 contexts)
302
  - **Recommendation:** Context-3 or Context-4 for text generation
303
 
304
  ---
 
314
 
315
  | Metric | Value |
316
  |--------|-------|
317
+ | Vocabulary Size | 271,198 |
318
+ | Total Tokens | 12,478,531 |
319
+ | Mean Frequency | 46.01 |
320
  | Median Frequency | 3 |
321
+ | Frequency Std Dev | 1145.11 |
322
 
323
  ### Most Common Words
324
 
325
  | Rank | Word | Frequency |
326
  |------|------|-----------|
327
+ | 1 | و | 284,031 |
328
+ | 2 | بیر | 169,280 |
329
+ | 3 | اینگیلیسجه | 149,737 |
330
+ | 4 | قایناقلار | 141,945 |
331
+ | 5 | the | 114,439 |
332
+ | 6 | تاریخینده | 92,079 |
333
+ | 7 | قایناق‌لار | 90,963 |
334
+ | 8 | ایلده | 83,679 |
335
+ | 9 | شهرلری | 81,894 |
336
+ | 10 | طرفیندن | 80,132 |
337
 
338
  ### Least Common Words (from vocabulary)
339
 
340
  | Rank | Word | Frequency |
341
  |------|------|-----------|
342
+ | 1 | کساسیاسی | 2 |
343
+ | 2 | کالابری�� | 2 |
344
+ | 3 | کونتینوا | 2 |
345
+ | 4 | تحقیق‌لری | 2 |
346
+ | 5 | romanzo | 2 |
347
+ | 6 | strage | 2 |
348
+ | 7 | سۆره‌جینده | 2 |
349
+ | 8 | ایلکه‌لر | 2 |
350
+ | 9 | لائیکلیک | 2 |
351
+ | 10 | شاسکوه | 2 |
352
 
353
  ### Zipf's Law Analysis
354
 
355
  | Metric | Value |
356
  |--------|-------|
357
+ | Zipf Coefficient | 1.1609 |
358
+ | R² (Goodness of Fit) | 0.995521 |
359
  | Adherence Quality | **excellent** |
360
 
361
  ### Coverage Analysis
362
 
363
  | Top N Words | Coverage |
364
  |-------------|----------|
365
+ | Top 100 | 34.4% |
366
+ | Top 1,000 | 64.8% |
367
+ | Top 5,000 | 79.6% |
368
+ | Top 10,000 | 84.6% |
369
 
370
  ### Key Findings
371
 
372
+ - **Zipf Compliance:** R²=0.9955 indicates excellent adherence to Zipf's law
373
+ - **High Frequency Dominance:** Top 100 words cover 34.4% of corpus
374
+ - **Long Tail:** 261,198 words needed for remaining 15.4% coverage
375
 
376
  ---
377
  ## 5. Word Embeddings Evaluation
 
384
 
385
  ![t-SNE Sentences](visualizations/tsne_sentences.png)
386
 
 
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 | 0.8282 🏆 | 0.3614 | N/A | N/A |
398
+ | **mono_64d** | 64 | 0.7952 | 0.3099 | N/A | N/A |
399
+ | **mono_128d** | 128 | 0.7570 | 0.2493 | N/A | N/A |
400
 
401
  ### Key Findings
402
 
403
+ - **Best Isotropy:** mono_32d with 0.8282 (more uniform distribution)
404
+ - **Semantic Density:** Average pairwise similarity of 0.3069. 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
+
424
+ 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.
425
+
426
+ #### Productive Prefixes
427
+ | Prefix | Examples |
428
+ |--------|----------|
429
+
430
+ #### Productive Suffixes
431
+ | Suffix | Examples |
432
+ |--------|----------|
433
+ | `-ین` | ائتدیگی‌نین, دالینین, لشکرینین |
434
+ | `-ان` | سيران, کاپیتان, تاپیلمایان |
435
+
436
+ ### 6.3 Bound Stems (Lexical Roots)
437
+
438
+ 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.
439
+
440
+ | Stem | Cohesion | Substitutability | Examples |
441
+ |------|----------|------------------|----------|
442
+ | `رلری` | 1.93x | 205 contexts | ارلری, یرلری, دیرلری |
443
+ | `اقلا` | 1.95x | 131 contexts | ناقلا, آیاقلا, آياقلا |
444
+ | `قلار` | 2.11x | 54 contexts | لیقلار, حقلاری, ماقلار |
445
+ | `تیند` | 1.93x | 72 contexts | تیندل, تینده, اتیندن |
446
+ | `یبدی` | 2.32x | 31 contexts | آلیبدی, ییبدیر, گلیبدی |
447
+ | `اریخ` | 1.93x | 41 contexts | تا��یخ, تاریخه, ‌تاریخ |
448
+ | `ولوب` | 1.70x | 60 contexts | کولوب, گولوب, سولوب |
449
+ | `ئرلش` | 2.00x | 24 contexts | يئرلشن, یئرلشن, یئرلشه |
450
+ | `ریخی` | 2.03x | 22 contexts | مریخی, ریخین, تاریخی |
451
+ | `یناق` | 1.87x | 27 contexts | سیناق, قیناق, ایناق |
452
+ | `هرلر` | 2.14x | 17 contexts | شهرلر, شهرلره, شهرلري |
453
+ | `یلیس` | 1.56x | 43 contexts | هیلیس, یلیسی, تیلیس |
454
+
455
+ ### 6.4 Affix Compatibility (Co-occurrence)
456
+
457
+ This table shows which prefixes and suffixes most frequently co-occur on the same stems, revealing the 'stacking' rules of the language's morphology.
458
+
459
+ *No significant affix co-occurrences detected.*
460
+
461
+
462
+ ### 6.5 Recursive Morpheme Segmentation
463
+
464
+ Using **Recursive Hierarchical Substitutability**, we decompose complex words into their constituent morphemes. This approach handles nested affixes (e.g., `prefix-prefix-root-suffix`).
465
+
466
+ | Word | Suggested Split | Confidence | Stem |
467
+ |------|-----------------|------------|------|
468
+ | وطنداشلارینین | **`وطنداشلار-ین-ین`** | 6.0 | `وطنداشلار` |
469
+ | تورپاق‌لارینین | **`تورپاق‌لار-ین-ین`** | 6.0 | `تورپاق‌لار` |
470
+ | دئموکرات‌لارینین | **`دئموکرات‌لار-ین-ین`** | 6.0 | `دئموکرات‌لار` |
471
+ | اوستانلارینان | **`اوستانلار-ین-ان`** | 6.0 | `اوستانلار` |
472
+ | تولیدینین | **`تولید-ین-ین`** | 6.0 | `تولید` |
473
+ | المنتلرینین | **`المنتلر-ین-ین`** | 6.0 | `المنتلر` |
474
+ | جومهوریتلرین | **`جومهوریتلر-ین`** | 4.5 | `جومهوریتلر` |
475
+ | سیمالارین | **`سیمالار-ین`** | 4.5 | `سیمالار` |
476
+ | ژوآن‌لارین | **`ژوآن‌لار-ین`** | 4.5 | `ژوآن‌لار` |
477
+ | ماشین‌لارین | **`ماشین‌لار-ین`** | 4.5 | `ماشین‌لار` |
478
+ | بدوی‌لرین | **`بدوی‌لر-ین`** | 4.5 | `بدوی‌لر` |
479
+ | تبریزلیلرین | **`تبریزلیلر-ین`** | 4.5 | `تبریزلیلر` |
480
+ | ناخوشلوقلارین | **`ناخوشلوقلار-ین`** | 4.5 | `ناخوشلوقلار` |
481
+ | اویکونیمین | **`اویکونیم-ین`** | 4.5 | `اویکونیم` |
482
+ | ائرککلرین | **`ائرککلر-ین`** | 4.5 | `ائرککلر` |
483
+
484
+ ### 6.6 Linguistic Interpretation
485
+
486
+ > **Automated Insight:**
487
+ The language AZB 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.
488
 
489
  ---
490
+ ## 7. Summary & Recommendations
491
 
492
  ![Performance Dashboard](visualizations/performance_dashboard.png)
493
 
 
495
 
496
  | Component | Recommended | Rationale |
497
  |-----------|-------------|-----------|
498
+ | Tokenizer | **64k BPE** | Best compression (4.15x) |
499
+ | N-gram | **2-gram** | Lowest perplexity (528) |
500
+ | Markov | **Context-4** | Highest predictability (96.6%) |
501
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
502
 
503
+
504
  ---
505
  ## Appendix: Metrics Glossary & Interpretation Guide
506
 
 
690
  author = {Kamali, Omar},
691
  title = {Wikilangs: Open NLP Models for Wikipedia Languages},
692
  year = {2025},
693
+ doi = {10.5281/zenodo.18073153},
694
+ publisher = {Zenodo},
695
  url = {https://huggingface.co/wikilangs}
696
  institution = {Omneity Labs}
697
  }
 
707
  - 🤗 Models: [huggingface.co/wikilangs](https://huggingface.co/wikilangs)
708
  - 📊 Data: [wikipedia-monthly](https://huggingface.co/datasets/omarkamali/wikipedia-monthly)
709
  - 👤 Author: [Omar Kamali](https://huggingface.co/omarkamali)
710
+ - 🤝 Sponsor: [Featherless AI](https://featherless.ai)
711
  ---
712
  *Generated by Wikilangs Models Pipeline*
713
 
714
+ *Report Date: 2026-01-03 06:14:53*
models/embeddings/monolingual/azb_128d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e36e34c4aa5231270647ed20effcf7427bf66e34bf81f236183cc4d2910eef2d
3
- size 1164785458
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0affe475b2c95d7dbb76d561e1c19235d6d09956c11cb77486d2ed42c5bc91f8
3
+ size 1145287401
models/embeddings/monolingual/azb_128d_metadata.json CHANGED
@@ -3,11 +3,13 @@
3
  "dimension": 128,
4
  "version": "monolingual",
5
  "training_params": {
6
- "dim": 128,
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
- "epochs": 5
 
 
11
  },
12
- "vocab_size": 134712
13
  }
 
3
  "dimension": 128,
4
  "version": "monolingual",
5
  "training_params": {
6
+ "algorithm": "skipgram",
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
+ "epochs": 5,
11
+ "encoding_method": "rope",
12
+ "dim": 128
13
  },
14
+ "vocab_size": 116118
15
  }
models/embeddings/monolingual/azb_32d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:24dbd964d896595164ff03aa344de312318ab9024fadf1dda4173202ae37885b
3
- size 293326642
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd0d8e4ac83c586e8f75145c0ce099bef44ff8d39e955040bc9454b60e32022c
3
+ size 288108777
models/embeddings/monolingual/azb_32d_metadata.json CHANGED
@@ -3,11 +3,13 @@
3
  "dimension": 32,
4
  "version": "monolingual",
5
  "training_params": {
6
- "dim": 32,
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
- "epochs": 5
 
 
11
  },
12
- "vocab_size": 134712
13
  }
 
3
  "dimension": 32,
4
  "version": "monolingual",
5
  "training_params": {
6
+ "algorithm": "skipgram",
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
+ "epochs": 5,
11
+ "encoding_method": "rope",
12
+ "dim": 32
13
  },
14
+ "vocab_size": 116118
15
  }
models/embeddings/monolingual/azb_64d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6193385dacda8dfbeee9fb60f5514bc20a74571e0a03c1d6372443118dd7271a
3
- size 583812914
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2aae88ec4bdf46e7654ecacada31d2b3be0690719379b657e44df7793b45fa65
3
+ size 573834985
models/embeddings/monolingual/azb_64d_metadata.json CHANGED
@@ -3,11 +3,13 @@
3
  "dimension": 64,
4
  "version": "monolingual",
5
  "training_params": {
6
- "dim": 64,
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
- "epochs": 5
 
 
11
  },
12
- "vocab_size": 134712
13
  }
 
3
  "dimension": 64,
4
  "version": "monolingual",
5
  "training_params": {
6
+ "algorithm": "skipgram",
7
  "min_count": 5,
8
  "window": 5,
9
  "negative": 5,
10
+ "epochs": 5,
11
+ "encoding_method": "rope",
12
+ "dim": 64
13
  },
14
+ "vocab_size": 116118
15
  }
models/subword_markov/azb_markov_ctx1_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b5b4c16c0cc262c904b79b490c2cb1ed8fa1b624111ee815c8a153b6470de540
3
- size 249202
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ff3da37abf00dd52bdcb28478a692c4366754cc1a111aa42559670cd7ea3aee
3
+ size 247087
models/subword_markov/azb_markov_ctx1_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_contexts": 2934,
6
- "total_transitions": 117347430
7
  }
 
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_contexts": 3409,
6
+ "total_transitions": 91613484
7
  }
models/subword_markov/azb_markov_ctx2_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1bb9b0eb974612ded94c2843ab693fd0d8bae85b9e9719cc399abf8806f0d8e6
3
- size 2049903
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1090e7d0037ce29d3cddcfe27f87428fcd5c872805dcf5e3d4202b9f83929b66
3
+ size 1650829
models/subword_markov/azb_markov_ctx2_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_contexts": 31498,
6
- "total_transitions": 117102967
7
  }
 
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_contexts": 30905,
6
+ "total_transitions": 91370829
7
  }
models/subword_markov/azb_markov_ctx3_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:364a895a59b1416e468680327f44afdb5268c191c6d191e6b325691c0f3c3920
3
- size 10556130
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:06a5083f143c1f0688e95d617ee8d409f5c78a44f6ea6c79e014253dcf249ac5
3
+ size 7956042
models/subword_markov/azb_markov_ctx3_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_contexts": 262151,
6
- "total_transitions": 116858504
7
  }
 
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_contexts": 203056,
6
+ "total_transitions": 91128174
7
  }
models/subword_markov/azb_markov_ctx4_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6bbc3fbabaaf693d43117cfaccb8266ae07186aa1c883be3dd208a2f9aad8724
3
- size 35507016
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4bd1bdcfb24b934a84145355781bb3e8cf414b875f467fc28d048b5d64ec5d1
3
+ size 26328560
models/subword_markov/azb_markov_ctx4_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_contexts": 1333406,
6
- "total_transitions": 116614041
7
  }
 
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_contexts": 953883,
6
+ "total_transitions": 90885519
7
  }
models/subword_ngram/azb_2gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ccea8a8d09e5e2c50495929c3ed6209d6250e502dbf821570083b812edd44ab9
3
- size 198369
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:383d8448d14aee343af41336d2c34807c417414896fb0905958afee3cec1c929
3
+ size 182774
models/subword_ngram/azb_2gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_ngrams": 14131,
6
- "total_ngrams": 117347430
7
  }
 
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_ngrams": 12648,
6
+ "total_ngrams": 91613484
7
  }
models/subword_ngram/azb_3gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1e3575791a79055ab8760da4b584e9186ebd76add76ef9131f8b1c690f6f619c
3
- size 1733384
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed1d402ce0eac47e37053808b7981cf3eb00b76c5e2e5e887890b036ff01d869
3
+ size 1377099
models/subword_ngram/azb_3gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_ngrams": 141425,
6
- "total_ngrams": 117102967
7
  }
 
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_ngrams": 106644,
6
+ "total_ngrams": 91370829
7
  }
models/subword_ngram/azb_4gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a2d7bf39c2289a572f7d6d41c6ee5b70e899bc025d9c7d73b34bf4e38cac70f4
3
- size 9977940
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cde924d7fc46a3c6a14d25e38135fbf37a14e774606b67187d527dd381b1d75f
3
+ size 7368829
models/subword_ngram/azb_4gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "azb",
5
- "unique_ngrams": 804296,
6
- "total_ngrams": 116858504
7
  }
 
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "azb",
5
+ "unique_ngrams": 581225,
6
+ "total_ngrams": 91128174
7
  }
models/tokenizer/azb_tokenizer_16k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3884a029fd3957ce39121cce73cb5a8166d611ce94cef5675b43779699c0bd8a
3
- size 524898
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4f3786bae820ff2d6cc627dbdc96f1ede4586a6afa95391543c078559a7c125
3
+ size 527370
models/tokenizer/azb_tokenizer_16k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/azb_tokenizer_32k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7da271a7e77b591daaf850d5e0ca617ac859c9d9797bfc37fd65791b5cfc15fe
3
- size 819972
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5380be73c3e18000f3f24bd0d0309c1b6d15259c73623f50c42cea3c6f8992d
3
+ size 820944
models/tokenizer/azb_tokenizer_32k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/azb_tokenizer_64k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:54cad8dd709976bcc93887547846baf55cbc48f7161ddab74a438fd09acb2685
3
- size 1430728
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:efb9e4835aafb1f0f10715a303c2ac20f1f169ed4ffc46f49d685eea42a8bfae
3
+ size 1430567
models/tokenizer/azb_tokenizer_64k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/azb_tokenizer_8k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2ecb3b8a28e5f01ac0a0633151869481ce9b8a3af8ad9c54007045e09589ce83
3
- size 382716
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6b1776330fc0caaa7cc96df48505e83e46f3a00daa53e66e6a94354397d3dd8
3
+ size 384993
models/tokenizer/azb_tokenizer_8k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/vocabulary/azb_vocabulary.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4c0ed41ec149e3751f6a1ace2372fcee3d368b53c6355b2765ae7b98caed8924
3
- size 4872673
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1346fa231de1649149828cb0cae257f644b1914797c84a5a8c4b56c62e716253
3
+ size 4348211
models/vocabulary/azb_vocabulary_metadata.json CHANGED
@@ -1,16 +1,17 @@
1
  {
2
  "language": "azb",
3
- "vocabulary_size": 317640,
 
4
  "statistics": {
5
- "type_token_ratio": 0.05786680977434657,
6
  "coverage": {
7
- "top_100": 0.35375568505036664,
8
- "top_1000": 0.6410291673928461,
9
- "top_5000": 0.7797111502624887,
10
- "top_10000": 0.8250270483868463
11
  },
12
- "hapax_count": 688404,
13
- "hapax_ratio": 0.6842682825005666,
14
- "total_documents": 244463
15
  }
16
  }
 
1
  {
2
  "language": "azb",
3
+ "vocabulary_size": 271198,
4
+ "variant": "full",
5
  "statistics": {
6
+ "type_token_ratio": 0.05623983023140009,
7
  "coverage": {
8
+ "top_100": 0.33209927062556827,
9
+ "top_1000": 0.6250560987377987,
10
+ "top_5000": 0.7676335196346162,
11
+ "top_10000": 0.8163255618590586
12
  },
13
+ "hapax_count": 456252,
14
+ "hapax_ratio": 0.627193621554746,
15
+ "total_documents": 242655
16
  }
17
  }
models/word_markov/azb_markov_ctx1_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0f8d5ae87507c80aafd715379fb38b0179c5049568d6f9e2b033dffefdfb0395
3
- size 49308871
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:816eaec35f734257259dcbb9d574a8c019635474f0de8b0daca5e593a77ac78b
3
+ size 42654567
models/word_markov/azb_markov_ctx1_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_contexts": 1005885,
6
- "total_transitions": 23557605
7
  }
 
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_contexts": 726930,
6
+ "total_transitions": 12692128
7
  }
models/word_markov/azb_markov_ctx2_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4c8a73136462b690d7c5514e30b6f6b274d36f9f121fbafeb5e99c8af7ec34c3
3
- size 115872504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46d9babbdaaaad5f5dc9ac4a05883d1447bd376822b6b3dc01d9bfa81167154f
3
+ size 101955475
models/word_markov/azb_markov_ctx2_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_contexts": 4302287,
6
- "total_transitions": 23313143
7
  }
 
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_contexts": 3693091,
6
+ "total_transitions": 12449473
7
  }
models/word_markov/azb_markov_ctx3_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:acd2b58441a2312fe6aff4e924e858f5d85a5843063ac2adf427f032e6628d22
3
- size 173864269
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:950723c3851972daff338e7c32d25b6b2711a549f145cfc4f36896cd3c7ac632
3
+ size 137860035
models/word_markov/azb_markov_ctx3_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 3,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_contexts": 7369809,
6
- "total_transitions": 23068685
7
  }
 
2
  "context_size": 3,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_contexts": 5447170,
6
+ "total_transitions": 12206818
7
  }
models/word_markov/azb_markov_ctx4_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:849a5c01ecae01be263fc6446cd008e4ee8cd861a0e83cd8e5da46a5f178d5ae
3
- size 222555153
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a8542b69e6498bf99f08ac81a2c69e6e10cfaec85f6041228b2d33e47374fa2f
3
+ size 167254340
models/word_markov/azb_markov_ctx4_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 4,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_contexts": 9387722,
6
- "total_transitions": 22824244
7
  }
 
2
  "context_size": 4,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_contexts": 6178524,
6
+ "total_transitions": 11964164
7
  }
models/word_ngram/azb_2gram_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5d9b9580f5793b668288ab3a7773338440486cda750c383b557c68ac0bc9221f
3
- size 4287214
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:38d788975186225cbd44f6dc3f5933b5ef0372dd125cccbf762b7c146ac00818
3
+ size 2898203
models/word_ngram/azb_2gram_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 2,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_ngrams": 268529,
6
- "total_ngrams": 23557605
7
  }
 
2
  "n": 2,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_ngrams": 158908,
6
+ "total_ngrams": 12692128
7
  }
models/word_ngram/azb_3gram_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ec70d427e623958a19690dad68f77970e4b6ff77368e7285059f2db52e46e7e6
3
- size 9355723
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e3b2c502cac90ffea868839ff83a17e6f50d9e2b94a2406bbe1479ad2b587f88
3
+ size 4982020
models/word_ngram/azb_3gram_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 3,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_ngrams": 512828,
6
- "total_ngrams": 23313143
7
  }
 
2
  "n": 3,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_ngrams": 236749,
6
+ "total_ngrams": 12449473
7
  }
models/word_ngram/azb_4gram_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:6b30786e6c43d7ce3c871df1db1b870820ce26c16155e9f336633c143d3099b5
3
- size 19762130
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ca451d9d2e6c1c4cd584e9ce07d419f547d1fe4ce2dc69563f37268bba623e4e
3
+ size 10534721
models/word_ngram/azb_4gram_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 4,
3
  "variant": "word",
4
  "language": "azb",
5
- "unique_ngrams": 979944,
6
- "total_ngrams": 23068685
7
  }
 
2
  "n": 4,
3
  "variant": "word",
4
  "language": "azb",
5
+ "unique_ngrams": 426395,
6
+ "total_ngrams": 12206818
7
  }
visualizations/embedding_isotropy.png CHANGED
visualizations/embedding_norms.png CHANGED
visualizations/embedding_similarity.png CHANGED

Git LFS Details

  • SHA256: ab24ca2b3b79760a39cd0440340e66fb767190adb057a0f9d67be331ba0b6efb
  • Pointer size: 131 Bytes
  • Size of remote file: 159 kB

Git LFS Details

  • SHA256: 2460104e6a6d554b06d36f1014d3bc41ed179423cf5fbb4c0307247fd2b45755
  • Pointer size: 131 Bytes
  • Size of remote file: 167 kB
visualizations/markov_branching.png CHANGED
visualizations/markov_contexts.png CHANGED