omarkamali commited on
Commit
8444673
·
verified ·
1 Parent(s): 57165bc

Upload all models and assets for bh (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 +188 -151
  3. models/embeddings/aligned/bh_128d.bin +3 -0
  4. models/embeddings/aligned/bh_128d.meta.json +1 -0
  5. models/embeddings/aligned/bh_128d.projection.npy +3 -0
  6. models/embeddings/aligned/bh_128d_metadata.json +8 -0
  7. models/embeddings/aligned/bh_32d.bin +3 -0
  8. models/embeddings/aligned/bh_32d.meta.json +1 -0
  9. models/embeddings/aligned/bh_32d.projection.npy +3 -0
  10. models/embeddings/aligned/bh_32d_metadata.json +8 -0
  11. models/embeddings/aligned/bh_64d.bin +3 -0
  12. models/embeddings/aligned/bh_64d.meta.json +1 -0
  13. models/embeddings/aligned/bh_64d.projection.npy +3 -0
  14. models/embeddings/aligned/bh_64d_metadata.json +8 -0
  15. models/embeddings/monolingual/bh_128d.bin +2 -2
  16. models/embeddings/monolingual/bh_128d_metadata.json +1 -1
  17. models/embeddings/monolingual/bh_32d.bin +2 -2
  18. models/embeddings/monolingual/bh_32d_metadata.json +1 -1
  19. models/embeddings/monolingual/bh_64d.bin +2 -2
  20. models/embeddings/monolingual/bh_64d_metadata.json +1 -1
  21. models/subword_markov/bh_markov_ctx1_subword.parquet +2 -2
  22. models/subword_markov/bh_markov_ctx1_subword_metadata.json +2 -2
  23. models/subword_markov/bh_markov_ctx2_subword.parquet +2 -2
  24. models/subword_markov/bh_markov_ctx2_subword_metadata.json +2 -2
  25. models/subword_markov/bh_markov_ctx3_subword.parquet +2 -2
  26. models/subword_markov/bh_markov_ctx3_subword_metadata.json +2 -2
  27. models/subword_markov/bh_markov_ctx4_subword.parquet +2 -2
  28. models/subword_markov/bh_markov_ctx4_subword_metadata.json +2 -2
  29. models/subword_ngram/bh_2gram_subword.parquet +2 -2
  30. models/subword_ngram/bh_2gram_subword_metadata.json +2 -2
  31. models/subword_ngram/bh_3gram_subword.parquet +2 -2
  32. models/subword_ngram/bh_3gram_subword_metadata.json +2 -2
  33. models/subword_ngram/bh_4gram_subword.parquet +2 -2
  34. models/subword_ngram/bh_4gram_subword_metadata.json +2 -2
  35. models/subword_ngram/bh_5gram_subword.parquet +3 -0
  36. models/subword_ngram/bh_5gram_subword_metadata.json +7 -0
  37. models/tokenizer/bh_tokenizer_16k.model +2 -2
  38. models/tokenizer/bh_tokenizer_16k.vocab +0 -0
  39. models/tokenizer/bh_tokenizer_32k.model +2 -2
  40. models/tokenizer/bh_tokenizer_32k.vocab +0 -0
  41. models/tokenizer/bh_tokenizer_64k.model +2 -2
  42. models/tokenizer/bh_tokenizer_64k.vocab +0 -0
  43. models/tokenizer/bh_tokenizer_8k.model +2 -2
  44. models/tokenizer/bh_tokenizer_8k.vocab +0 -0
  45. models/vocabulary/bh_vocabulary.parquet +2 -2
  46. models/vocabulary/bh_vocabulary_metadata.json +9 -9
  47. models/word_markov/bh_markov_ctx1_word.parquet +2 -2
  48. models/word_markov/bh_markov_ctx1_word_metadata.json +2 -2
  49. models/word_markov/bh_markov_ctx2_word.parquet +2 -2
  50. models/word_markov/bh_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: bh
3
- language_name: BH
4
  language_family: indoaryan_central
5
  tags:
6
  - wikilangs
@@ -10,11 +10,21 @@ tags:
10
  - n-gram
11
  - markov
12
  - wikipedia
 
 
 
 
 
 
 
 
 
 
13
  - monolingual
14
  - family-indoaryan_central
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: 4.103
27
  - name: best_isotropy
28
  type: isotropy
29
- value: 0.8668
30
  - name: vocabulary_size
31
  type: vocab
32
  value: 0
33
  generated: 2026-01-03
34
  ---
35
 
36
- # BH - Wikilangs Models
37
  ## Comprehensive Research Report & Full Ablation Study
38
 
39
- This repository contains NLP models trained and evaluated by Wikilangs, specifically on **BH** 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)
@@ -80,47 +90,47 @@ 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** | 3.436x | 3.44 | 0.1753% | 369,577 |
84
- | **16k** | 3.741x | 3.74 | 0.1909% | 339,439 |
85
- | **32k** | 3.960x | 3.96 | 0.2021% | 320,666 |
86
- | **64k** | 4.103x 🏆 | 4.11 | 0.2094% | 309,485 |
87
 
88
  ### Tokenization Examples
89
 
90
  Below are sample sentences tokenized with each vocabulary size:
91
 
92
- **Sample 1:** `अँगारपाथर भारत के झारखंड राज्य में धनबाद शहर के एगो मोहल्ला बाटे। संदर्भ के शहर‏...`
93
 
94
  | Vocab | Tokens | Count |
95
  |-------|--------|-------|
96
- | 8k | `▁अँ गार पा थर ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁धन ... (+12 more)` | 22 |
97
- | 16k | `▁अँ गार पा थर ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁धनबाद ... (+11 more)` | 21 |
98
- | 32k | `▁अँ गार पाथर ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁धनबाद ▁शहर ... (+10 more)` | 20 |
99
- | 64k | `▁अँगारपाथर ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁धनबाद ▁शहर ▁के ▁एगो ... (+8 more)` | 18 |
100
 
101
- **Sample 2:** `जून ग्रेगरियन कैलेंडर के छठवाँ महीना ह। घटना तिहुआर अउरी दूसर महत्व के दिन अउरी ...`
102
 
103
  | Vocab | Tokens | Count |
104
  |-------|--------|-------|
105
- | 8k | `▁जून ▁ग्रेगरियन ▁कैलेंडर ▁के ▁छठ वाँ ▁महीना ▁ह ▁घटना ... (+14 more)` | 24 |
106
- | 16k | `▁जून ▁ग्रेगरियन ▁कैलेंडर ▁के ▁छठवाँ ▁महीना ▁ह ▁घटना ▁तिहुआर ... (+13 more)` | 23 |
107
- | 32k | `▁जून ▁ग्रेगरियन ▁कैलेंडर ▁के ▁छठवाँ ▁महीना ▁ह ▁घटना ▁तिहुआर ... (+13 more)` | 23 |
108
- | 64k | `▁जून ▁ग्रेगरियन ▁कैलेंडर ▁के ▁छठवाँ ▁महीना ▁ह ▁घटना ▁तिहुआर ... (+13 more)` | 23 |
109
 
110
- **Sample 3:** `बदायूँ जिला उत्तर प्रदेश की बरेली मंडल में एगो जिला बाटे जौना के मुख्यालय बदायूँ...`
111
 
112
  | Vocab | Tokens | Count |
113
  |-------|--------|-------|
114
- | 8k | `▁ब दाय ूँ ▁जिला ▁उत्तर ▁प्रदेश ▁की ▁ब रेली ▁मंडल ... (+19 more)` | 29 |
115
- | 16k | `▁ब दाय ूँ ▁जिला ▁उत्तर ▁प्रदेश ▁की ▁बरेली ▁मंडल ▁में ... (+18 more)` | 28 |
116
- | 32k | `▁ब दायूँ ▁जिला ▁उत्तर ▁प्रदेश ▁की ▁बरेली ▁मंडल ▁में ▁एगो ... (+16 more)` | 26 |
117
- | 64k | `▁बदायूँ ▁जिला ▁उत्तर ▁प्रदेश ▁की ▁बरेली ▁मंडल ▁में ▁एगो ▁जिला ... (+14 more)` | 24 |
118
 
119
 
120
  ### Key Findings
121
 
122
- - **Best Compression:** 64k achieves 4.103x compression
123
- - **Lowest UNK Rate:** 8k with 0.1753% 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 | 9,163 | 13.16 | 29,857 | 16.5% | 43.3% |
141
- | **2-gram** | Subword | 1,500 🏆 | 10.55 | 21,796 | 39.5% | 76.5% |
142
- | **3-gram** | Word | 13,824 | 13.75 | 38,729 | 15.8% | 36.1% |
143
- | **3-gram** | Subword | 11,162 | 13.45 | 93,652 | 16.7% | 42.2% |
144
- | **4-gram** | Word | 17,666 | 14.11 | 53,247 | 17.6% | 35.4% |
145
- | **4-gram** | Subword | 44,915 | 15.45 | 295,453 | 9.1% | 27.7% |
 
 
146
 
147
  ### Top 5 N-grams by Size
148
 
@@ -150,20 +162,20 @@ Below are sample sentences tokenized with each vocabulary size:
150
 
151
  | Rank | N-gram | Count |
152
  |------|--------|-------|
153
- | 1 | `सभ के` | 4,161 |
154
- | 2 | `भारत के` | 3,814 |
155
- | 3 | `रूप में` | 3,157 |
156
- | 4 | `के रूप` | 2,933 |
157
- | 5 | `देखल जाय` | 2,149 |
158
 
159
  **3-grams (Word):**
160
 
161
  | Rank | N-gram | Count |
162
  |------|--------|-------|
163
- | 1 | `के रूप में` | 2,740 |
164
- | 2 | `इहो देखल जाय` | 2,002 |
165
- | 3 | `के हिसाब से` | 1,423 |
166
- | 4 | `संदर्भ बाहरी कड़ी` | 1,392 |
167
  | 5 | `शहर आ कस्बा` | 1,209 |
168
 
169
  **4-grams (Word):**
@@ -171,47 +183,67 @@ Below are sample sentences tokenized with each vocabulary size:
171
  | Rank | N-gram | Count |
172
  |------|--------|-------|
173
  | 1 | `के शहर आ कस्बा` | 1,206 |
174
- | 2 | `बाटे इहो देखल जाय` | 780 |
175
- | 3 | `राज्य में एक ठो` | 667 |
176
  | 4 | `के हिसाब से ई` | 539 |
177
  | 5 | `में एगो जिला बाटे` | 536 |
178
 
 
 
 
 
 
 
 
 
 
 
179
  **2-grams (Subword):**
180
 
181
  | Rank | N-gram | Count |
182
  |------|--------|-------|
183
- | 1 | `के _` | 114,253 |
184
- | 2 | `_ के` | 110,824 |
185
- | 3 | `र _` | 75,001 |
186
- | 4 | `ल _` | 68,413 |
187
- | 5 | `न _` | 54,528 |
188
 
189
  **3-grams (Subword):**
190
 
191
  | Rank | N-gram | Count |
192
  |------|--------|-------|
193
- | 1 | `_ के _` | 109,027 |
194
- | 2 | `_ में _` | 44,490 |
195
- | 3 | `_ आ _` | 29,937 |
196
- | 4 | `_ से _` | 20,956 |
197
- | 5 | `ल _ जा` | 13,886 |
198
 
199
  **4-grams (Subword):**
200
 
201
  | Rank | N-gram | Count |
202
  |------|--------|-------|
203
- | 1 | `न _ के _` | 9,495 |
204
- | 2 | `_ स भ _` | 8,569 |
205
- | 3 | `_ ए गो _` | 8,113 |
206
- | 4 | `र _ के _` | 7,353 |
207
- | 5 | `ल _ जा ला` | 7,231 |
 
 
 
 
 
 
 
 
 
 
208
 
209
 
210
  ### Key Findings
211
 
212
- - **Best Perplexity:** 2-gram (subword) with 1,500
213
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
214
- - **Coverage:** Top-1000 patterns cover ~28% 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 | 0.8766 | 1.836 | 6.16 | 84,482 | 12.3% |
231
- | **1** | Subword | 0.9997 | 2.000 | 12.30 | 4,952 | 0.0% |
232
- | **2** | Word | 0.2946 | 1.227 | 1.77 | 519,009 | 70.5% |
233
- | **2** | Subword | 0.5586 | 1.473 | 4.02 | 60,879 | 44.1% |
234
- | **3** | Word | 0.1069 | 1.077 | 1.19 | 917,743 | 89.3% |
235
- | **3** | Subword | 0.5222 | 1.436 | 2.95 | 244,880 | 47.8% |
236
- | **4** | Word | 0.0351 🏆 | 1.025 | 1.05 | 1,088,288 | 96.5% |
237
- | **4** | Subword | 0.3351 | 1.261 | 1.87 | 721,221 | 66.5% |
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. `के कुछ अउरी पढ़े के इस्तेमाल होला नया विमानन के दैविक घटना जनम 11`
246
- 2. `में एकट्ठा क्षमता में राजा दशरथ एकर नक़ल उतारे जा सके ला बाद के`
247
- 3. `आ भोजपुरी में एक ठो इतिहासी भूबिज्ञान मिजोरम के नाँव ढेर बरफ के अबतक ले`
248
 
249
  **Context Size 2:**
250
 
251
- 1. `सभ के कक्षा ऑर्बिट सुरुज के सभसे पबित्र मानल जाला एह दिन के मतलब मैदान के मैदान`
252
- 2. `भारत के बारहवाँ कार्यकाल अनुसार 14वाँ वर्तमान में भारत के प्रतिनिधित्व यूरोपियन कमीशन द्वारा 12 मई`
253
- 3. `रूप में अवधारणा के अंतर्राष्ट्रीय बॉर्डर के रूप में निरूपण नक्शा कौनों इलाका के मिला लिहल जाय`
254
 
255
  **Context Size 3:**
256
 
257
- 1. `के रूप में बर्गीकरण कइल जाला कुछ दशा में घार्मिक कामकाज खातिर भी नगर के महतव बा`
258
- 2. `इहो देखल जाय बिहार सरकार बिहार के बिकास के रूप में एमबीए कइलें स्टैनफोर्ड में पढ़त घरी इनके`
259
- 3. `के हिसाब से दुनिया के छत्तीसवाँ देस हवे पूरबी हिस्सा में उत्तर से दक्खिन ओर फइलल बिसाल`
260
 
261
  **Context Size 4:**
262
 
263
- 1. `बाटे इहो देखल जाय ओडिशा के जिला भारत के जिला सभ के लिस्ट संदर्भ बाहरी कड़ी जिला समन्वय समिति`
264
- 2. `राज्य में एक ठो कसबा बाटे के शहर कस्बा जिला के शहर कस्बा बंगाल के शहर आ`
265
- 3. `के हिसाब से ई भारत के 191वाँ शहर बाटे जनगणना आँकड़ा के मोताबिक एह शहर में लिंगानुपात 992 आ`
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. `_क्रिके_की_दर्भ_रव_धन_हो`
275
- 2. `रण_योना_ल_मिल_बिकरेडिट`
276
- 3. `के_mainnaronamoxt`
277
 
278
  **Context Size 2:**
279
 
280
- 1. `के_पाँचवीं_सस्पेंसन_-टिप्पणी_के`
281
- 2. `_के_दिल्ली_गेम_बस_ce_th`
282
- 3. `र_स्थान_का_आउटवारा_से_सुन`
283
 
284
  **Context Size 3:**
285
 
286
- 1. `_के_राजा_हवे,_एक_ठो_छोट_`
287
- 2. `_में_इनहन_में_गलुन्गाम,_पीप`
288
- 3. `_आ_इनकर_दि_बीच_के_पहिला`
289
 
290
  **Context Size 4:**
291
 
292
- 1. `न_के_राजनर्तकी_अंबपाली_(आम्रपा`
293
- 2. `_सभ_में_राज_कर_सके_ला,_ब`
294
- 3. `_एगो_काल्पनिक_दुनिया_के_इस्तेमाल`
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 (721,221 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 | 38,858 |
318
- | Total Tokens | 1,245,419 |
319
- | Mean Frequency | 32.05 |
320
  | Median Frequency | 4 |
321
- | Frequency Std Dev | 665.90 |
322
 
323
  ### Most Common Words
324
 
325
  | Rank | Word | Frequency |
326
  |------|------|-----------|
327
- | 1 | के | 109,634 |
328
- | 2 | में | 46,202 |
329
- | 3 | आ | 30,024 |
330
- | 4 | से | 21,308 |
331
- | 5 | बा | 11,775 |
332
- | 6 | ई | 10,655 |
333
- | 7 | सभ | 8,830 |
334
- | 8 | बाटे | 8,519 |
335
- | 9 | एगो | 8,159 |
336
- | 10 | जाला | 8,051 |
337
 
338
  ### Least Common Words (from vocabulary)
339
 
340
  | Rank | Word | Frequency |
341
  |------|------|-----------|
342
- | 1 | मापे | 2 |
343
- | 2 | संयुक्‍त | 2 |
344
- | 3 | व्युत्क्रमानुपात | 2 |
345
- | 4 | सीमाएँ | 2 |
346
- | 5 | एहन | 2 |
347
- | 6 | परिपथ | 2 |
348
- | 7 | voltage | 2 |
349
- | 8 | विभवांतर | 2 |
350
- | 9 | वोल्ट | 2 |
351
- | 10 | एम्पियर | 2 |
352
 
353
  ### Zipf's Law Analysis
354
 
355
  | Metric | Value |
356
  |--------|-------|
357
- | Zipf Coefficient | 1.1200 |
358
- | R² (Goodness of Fit) | 0.994371 |
359
  | Adherence Quality | **excellent** |
360
 
361
  ### Coverage Analysis
362
 
363
  | Top N Words | Coverage |
364
  |-------------|----------|
365
- | Top 100 | 43.0% |
366
- | Top 1,000 | 69.5% |
367
  | Top 5,000 | 86.1% |
368
  | Top 10,000 | 91.7% |
369
 
370
  ### Key Findings
371
 
372
  - **Zipf Compliance:** R²=0.9944 indicates excellent adherence to Zipf's law
373
- - **High Frequency Dominance:** Top 100 words cover 43.0% of corpus
374
- - **Long Tail:** 28,858 words needed for remaining 8.3% 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 | 0.8668 🏆 | 0.3638 | N/A | N/A |
398
- | **mono_64d** | 64 | 0.8282 | 0.2819 | N/A | N/A |
399
- | **mono_128d** | 128 | 0.6394 | 0.2329 | N/A | N/A |
 
 
 
400
 
401
  ### Key Findings
402
 
403
- - **Best Isotropy:** mono_32d with 0.8668 (more uniform distribution)
404
- - **Semantic Density:** Average pairwise similarity of 0.2929. 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
 
@@ -432,18 +467,18 @@ Bound stems are high-frequency subword units that are semantically cohesive but
432
 
433
  | Stem | Cohesion | Substitutability | Examples |
434
  |------|----------|------------------|----------|
435
- | `ther` | 2.68x | 27 contexts | other, there, rather |
436
- | `ight` | 2.68x | 21 contexts | fight, right, light |
437
- | `tion` | 2.59x | 21 contexts | action, nation, motion |
438
- | `ount` | 2.65x | 15 contexts | count, mount, amount |
439
- | `atio` | 2.62x | 15 contexts | ratio, nation, nations |
440
- | `ctio` | 2.61x | 14 contexts | action, fiction, auction |
441
- | `ater` | 2.67x | 11 contexts | eater, water, later |
442
- | `stat` | 2.63x | 10 contexts | stato, state, stats |
443
- | `vers` | 2.52x | 11 contexts | verse, covers, rivers |
444
- | `rati` | 2.58x | 9 contexts | ratio, rating, bharati |
445
- | `ment` | 2.50x | 9 contexts | cement, ferment, element |
446
- | `ical` | 2.57x | 8 contexts | radical, musical, typical |
447
 
448
  ### 6.4 Affix Compatibility (Co-occurrence)
449
 
@@ -462,7 +497,9 @@ Using **Recursive Hierarchical Substitutability**, we decompose complex words in
462
  ### 6.6 Linguistic Interpretation
463
 
464
  > **Automated Insight:**
465
- The language BH 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.
 
 
466
 
467
  ---
468
  ## 7. Summary & Recommendations
@@ -474,7 +511,7 @@ The language BH appears to be more isolating or has a highly fixed vocabulary. W
474
  | Component | Recommended | Rationale |
475
  |-----------|-------------|-----------|
476
  | Tokenizer | **64k BPE** | Best compression (4.10x) |
477
- | N-gram | **2-gram** | Lowest perplexity (1,500) |
478
  | Markov | **Context-4** | Highest predictability (96.5%) |
479
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
480
 
@@ -689,4 +726,4 @@ MIT License - Free for academic and commercial use.
689
  ---
690
  *Generated by Wikilangs Models Pipeline*
691
 
692
- *Report Date: 2026-01-03 07:15:04*
 
1
  ---
2
  language: bh
3
+ language_name: Bihari languages
4
  language_family: indoaryan_central
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-indoaryan_central
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: 4.105
37
  - name: best_isotropy
38
  type: isotropy
39
+ value: 0.8673
40
  - name: vocabulary_size
41
  type: vocab
42
  value: 0
43
  generated: 2026-01-03
44
  ---
45
 
46
+ # Bihari languages - Wikilangs Models
47
  ## Comprehensive Research Report & Full Ablation Study
48
 
49
+ This repository contains NLP models trained and evaluated by Wikilangs, specifically on **Bihari languages** 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** | 3.440x | 3.44 | 0.1739% | 367,965 |
94
+ | **16k** | 3.744x | 3.75 | 0.1893% | 338,089 |
95
+ | **32k** | 3.961x | 3.96 | 0.2003% | 319,582 |
96
+ | **64k** | 4.105x 🏆 | 4.11 | 0.2075% | 308,421 |
97
 
98
  ### Tokenization Examples
99
 
100
  Below are sample sentences tokenized with each vocabulary size:
101
 
102
+ **Sample 1:** `नेल्सन मंडेला दक्खिन अफिरका के पहिला करिया राष्ट्रपति पहिला चुनल गइल राष्ट्रपत...`
103
 
104
  | Vocab | Tokens | Count |
105
  |-------|--------|-------|
106
+ | 8k | `▁ने ल् सन ▁मंड ेला ▁दक्खिन ▁अफिरका ▁के ▁पहिला ▁करिया ... (+9 more)` | 19 |
107
+ | 16k | `▁ने ल्सन ▁मंड ेला ▁दक्खिन ▁अफिरका ▁के ▁पहिला ▁करिया ▁राष्ट्रपति ... (+8 more)` | 18 |
108
+ | 32k | `▁नेल्सन ▁मंड ेला ▁दक्खिन ▁अफिरका ▁के ▁पहिला ▁करिया ▁राष्ट्रपति ▁आ ... (+7 more)` | 17 |
109
+ | 64k | `▁नेल्सन ▁मंड ेला ▁दक्खिन ▁अफिरका ▁के ▁पहिला ▁करिया ▁राष्ट्रपति ▁आ ... (+7 more)` | 17 |
110
 
111
+ **Sample 2:** `बबुआ कलां भारत के झारखंड राज्य में एक ठो कसबा बाटे। के शहर‏‎ कस्बा`
112
 
113
  | Vocab | Tokens | Count |
114
  |-------|--------|-------|
115
+ | 8k | `▁ब ुआ ▁कला ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ... (+9 more)` | 19 |
116
+ | 16k | `▁ब बुआ ▁कला ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁एक ... (+8 more)` | 18 |
117
+ | 32k | `▁ब बुआ ▁कलां ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁एक ▁ठो ... (+7 more)` | 17 |
118
+ | 64k | `▁बबुआ ▁कलां ▁भारत ▁के ▁झारखंड ▁राज्य ▁में ▁एक ▁ठो ▁कसबा ... (+6 more)` | 16 |
119
 
120
+ **Sample 3:** `घटना जनम - मन्मथनाथ गुप्त - भारतीय स्वतन्त्रता संग्राम एगो प्रमुख क्रान्तिकारी...`
121
 
122
  | Vocab | Tokens | Count |
123
  |-------|--------|-------|
124
+ | 8k | `▁घटना ▁जनम ▁- ▁म न् नाथ ▁गुप्त ▁- ... (+28 more)` | 38 |
125
+ | 16k | `▁घटना ▁जनम ▁- ▁म न् मथ नाथ ▁गुप्त ▁- ▁भारतीय ... (+26 more)` | 36 |
126
+ | 32k | `▁घटना ▁जनम ▁- ▁मन् मथ नाथ ▁गुप्त ▁- ▁भारतीय ▁स्वतन्त्रता ... (+21 more)` | 31 |
127
+ | 64k | `▁घटना ▁जनम ▁- ▁मन्मथनाथ ▁गुप्त ▁- ▁भारतीय ▁स्वतन्त्रता ▁संग्राम ▁क ... (+17 more)` | 27 |
128
 
129
 
130
  ### Key Findings
131
 
132
+ - **Best Compression:** 64k achieves 4.105x compression
133
+ - **Lowest UNK Rate:** 8k with 0.1739% 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 | 9,136 | 13.16 | 29,778 | 16.5% | 43.4% |
151
+ | **2-gram** | Subword | 1,496 🏆 | 10.55 | 21,749 | 39.6% | 76.5% |
152
+ | **3-gram** | Word | 13,783 | 13.75 | 38,633 | 15.8% | 36.1% |
153
+ | **3-gram** | Subword | 11,127 | 13.44 | 93,435 | 16.7% | 42.3% |
154
+ | **4-gram** | Word | 17,572 | 14.10 | 53,047 | 17.6% | 35.4% |
155
+ | **4-gram** | Subword | 44,731 | 15.45 | 294,486 | 9.1% | 27.8% |
156
+ | **5-gram** | Word | 8,139 | 12.99 | 30,163 | 24.3% | 46.7% |
157
+ | **5-gram** | Subword | 95,769 | 16.55 | 421,404 | 6.3% | 19.7% |
158
 
159
  ### Top 5 N-grams by Size
160
 
 
162
 
163
  | Rank | N-gram | Count |
164
  |------|--------|-------|
165
+ | 1 | `सभ के` | 4,152 |
166
+ | 2 | `भारत के` | 3,812 |
167
+ | 3 | `रूप में` | 3,160 |
168
+ | 4 | `के रूप` | 2,936 |
169
+ | 5 | `देखल जाय` | 2,147 |
170
 
171
  **3-grams (Word):**
172
 
173
  | Rank | N-gram | Count |
174
  |------|--------|-------|
175
+ | 1 | `के रूप में` | 2,742 |
176
+ | 2 | `इहो देखल जाय` | 2,001 |
177
+ | 3 | `के हिसाब से` | 1,425 |
178
+ | 4 | `संदर्भ बाहरी कड़ी` | 1,391 |
179
  | 5 | `शहर आ कस्बा` | 1,209 |
180
 
181
  **4-grams (Word):**
 
183
  | Rank | N-gram | Count |
184
  |------|--------|-------|
185
  | 1 | `के शहर आ कस्बा` | 1,206 |
186
+ | 2 | `बाटे इहो देखल जाय` | 781 |
187
+ | 3 | `राज्य में एक ठो` | 666 |
188
  | 4 | `के हिसाब से ई` | 539 |
189
  | 5 | `में एगो जिला बाटे` | 536 |
190
 
191
+ **5-grams (Word):**
192
+
193
+ | Rank | N-gram | Count |
194
+ |------|--------|-------|
195
+ | 1 | `संदर्भ के शहर आ कस्बा` | 496 |
196
+ | 2 | `के जनगणना के हिसाब से` | 496 |
197
+ | 3 | `में एगो जिला बाटे एकर` | 465 |
198
+ | 4 | `जनसंख्या साल के जनगणना के` | 449 |
199
+ | 5 | `साल के जनगणना के हिसाब` | 448 |
200
+
201
  **2-grams (Subword):**
202
 
203
  | Rank | N-gram | Count |
204
  |------|--------|-------|
205
+ | 1 | `के _` | 114,017 |
206
+ | 2 | `_ के` | 110,574 |
207
+ | 3 | `र _` | 75,090 |
208
+ | 4 | `ल _` | 68,378 |
209
+ | 5 | `न _` | 54,576 |
210
 
211
  **3-grams (Subword):**
212
 
213
  | Rank | N-gram | Count |
214
  |------|--------|-------|
215
+ | 1 | `_ के _` | 108,779 |
216
+ | 2 | `_ में _` | 44,499 |
217
+ | 3 | `_ आ _` | 30,014 |
218
+ | 4 | `_ से _` | 20,994 |
219
+ | 5 | `ल _ जा` | 13,915 |
220
 
221
  **4-grams (Subword):**
222
 
223
  | Rank | N-gram | Count |
224
  |------|--------|-------|
225
+ | 1 | `न _ के _` | 9,485 |
226
+ | 2 | `_ स भ _` | 8,539 |
227
+ | 3 | `_ ए गो _` | 8,025 |
228
+ | 4 | `र _ के _` | 7,333 |
229
+ | 5 | `ल _ जा ला` | 7,264 |
230
+
231
+ **5-grams (Subword):**
232
+
233
+ | Rank | N-gram | Count |
234
+ |------|--------|-------|
235
+ | 1 | `_ बा टे । _` | 5,947 |
236
+ | 2 | `_ भा र त _` | 5,876 |
237
+ | 3 | `_ सं द र्भ _` | 5,473 |
238
+ | 4 | `_ t h e _` | 4,933 |
239
+ | 5 | `ल _ ग इ ल` | 4,916 |
240
 
241
 
242
  ### Key Findings
243
 
244
+ - **Best Perplexity:** 2-gram (subword) with 1,496
245
  - **Entropy Trend:** Decreases with larger n-grams (more predictable)
246
+ - **Coverage:** Top-1000 patterns cover ~20% 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 | 0.8731 | 1.832 | 6.14 | 84,373 | 12.7% |
263
+ | **1** | Subword | 0.9992 | 1.999 | 12.29 | 4,950 | 0.1% |
264
+ | **2** | Word | 0.2948 | 1.227 | 1.78 | 516,874 | 70.5% |
265
+ | **2** | Subword | 0.5582 | 1.472 | 4.02 | 60,819 | 44.2% |
266
+ | **3** | Word | 0.1070 | 1.077 | 1.19 | 914,610 | 89.3% |
267
+ | **3** | Subword | 0.5218 | 1.436 | 2.94 | 244,457 | 47.8% |
268
+ | **4** | Word | 0.0352 🏆 | 1.025 | 1.05 | 1,084,862 | 96.5% |
269
+ | **4** | Subword | 0.3349 | 1.261 | 1.87 | 719,467 | 66.5% |
270
 
271
  ### Generated Text Samples (Word-based)
272
 
 
274
 
275
  **Context Size 1:**
276
 
277
+ 1. `के काम कइल जाला के जिला भारत के संतान लक्ष्मीदास जे पर्यावरणी मेडिकल कॉलेज दारोगा`
278
+ 2. `में भगवान शिव के होखे ला दुनों जाना जाता था जो में जमल पानी प्रदूषण कहल`
279
+ 3. `आ निर्वासित दुनों ओर ना कौनों सामान सभ के नाट्यमण्डली के संभाव्यता अध्‍ययन राइट्स ऑफ हिज`
280
 
281
  **Context Size 2:**
282
 
283
+ 1. `सभ के समर्थन वाली मीरा कुमार रहली कहल गइल सन में बेंजामिन फ्रैंकलिन के`
284
+ 2. `भारत के 27वाँ शहर बाटे जनगणना आँकड़ा के मोताबिक राजा पृथु के नाँव सैयद शफ़ीक़ हुसैन रहल`
285
+ 3. `रूप में रखल जाला 23 मार्च locks down over 100 and 1 450 m oromediterranean zone nemoral`
286
 
287
  **Context Size 3:**
288
 
289
+ 1. `के रूप में भी देखल जाला पोसल जाला इन्हन कई गो अवतार कमल फूल अतिरिक्त`
290
+ 2. `इहो देखल जाय नारियल पानी नारियल गरी संदर्भ पानी`
291
+ 3. `के हिसाब से भारत के 476वाँ शहर बाटे जनगणना आँकड़ा के मोताबिक एह शहर में लिंगानुपात 934`
292
 
293
  **Context Size 4:**
294
 
295
+ 1. `बाटे इहो देखल जाय भारत के शहर संदर्भ के शहर कस्बा के शहर कस्बा प्रदेश के शहर`
296
+ 2. `राज्य में एक ठो कसबा बाटे इहो देखल जाय ग��जरात के जिला संदर्भ बाहरी कड़ी ऑफिशियल वेबसाइट के जिला`
297
+ 3. `के हिसाब से ई भारत के 204वाँ शहर बाटे जनगणना आँकड़ा के मोताबिक एह शहर में लिंगानुपात 883 आ`
298
 
299
 
300
  ### Generated Text Samples (Subword-based)
 
303
 
304
  **Context Size 1:**
305
 
306
+ 1. `_शार_खाई__भूगो_स्थापत्र_`
307
+ 2. `र_के_बत_oudeasuña`
308
+ 3. `के_में_का_djoriid_नित`
309
 
310
  **Context Size 2:**
311
 
312
+ 1. `के_युवल_जाति_बा_जे_दुनों_ची`
313
+ 2. `_के_तुलसीदास_लोगन-पूर्व_में`
314
+ 3. `र_द्वारा_पूरा_लोग_के_रूप_में`
315
 
316
  **Context Size 3:**
317
 
318
+ 1. `_के_पुरान_खान,_तसही_संभव`
319
+ 2. `_में_तीन_गो_देस_बनल_ईस्ट_`
320
+ 3. `_आ_सन्देश_पर_करा_जरूरत_`
321
 
322
  **Context Size 4:**
323
 
324
+ 1. `न_के_सीखल__एह_मंदिर,_मा`
325
+ 2. `_सभ_के_बिसेसता_के_कारण_मूल्य`
326
+ 3. `_एगो_नागरिक_उत्पादन_के_प्रति_`
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 (719,467 contexts)
334
  - **Recommendation:** Context-3 or Context-4 for text generation
335
 
336
  ---
 
346
 
347
  | Metric | Value |
348
  |--------|-------|
349
+ | Vocabulary Size | 38,630 |
350
+ | Total Tokens | 1,241,622 |
351
+ | Mean Frequency | 32.14 |
352
  | Median Frequency | 4 |
353
+ | Frequency Std Dev | 666.83 |
354
 
355
  ### Most Common Words
356
 
357
  | Rank | Word | Frequency |
358
  |------|------|-----------|
359
+ | 1 | के | 109,386 |
360
+ | 2 | में | 46,201 |
361
+ | 3 | आ | 30,101 |
362
+ | 4 | से | 21,341 |
363
+ | 5 | बा | 11,787 |
364
+ | 6 | ई | 10,672 |
365
+ | 7 | सभ | 8,798 |
366
+ | 8 | बाटे | 8,511 |
367
+ | 9 | जाला | 8,084 |
368
+ | 10 | एगो | 8,063 |
369
 
370
  ### Least Common Words (from vocabulary)
371
 
372
  | Rank | Word | Frequency |
373
  |------|------|-----------|
374
+ | 1 | बंटवारे | 2 |
375
+ | 2 | सीटेंराष्ट्रीय | 2 |
376
+ | 3 | पासवानभाकपा | 2 |
377
+ | 4 | शेयरमतदान | 2 |
378
+ | 5 | तिथिबहुमतराष्ट्रीय | 2 |
379
+ | 6 | गठबंधनमहागठबंधन | 2 |
380
+ | 7 | मैट्रिज़सितंबर | 2 |
381
+ | 8 | बोनो | 2 |
382
+ | 9 | नगद | 2 |
383
+ | 10 | रचनन | 2 |
384
 
385
  ### Zipf's Law Analysis
386
 
387
  | Metric | Value |
388
  |--------|-------|
389
+ | Zipf Coefficient | 1.1214 |
390
+ | R² (Goodness of Fit) | 0.994355 |
391
  | Adherence Quality | **excellent** |
392
 
393
  ### Coverage Analysis
394
 
395
  | Top N Words | Coverage |
396
  |-------------|----------|
397
+ | Top 100 | 43.1% |
398
+ | Top 1,000 | 69.6% |
399
  | Top 5,000 | 86.1% |
400
  | Top 10,000 | 91.7% |
401
 
402
  ### Key Findings
403
 
404
  - **Zipf Compliance:** R²=0.9944 indicates excellent adherence to Zipf's law
405
+ - **High Frequency Dominance:** Top 100 words cover 43.1% of corpus
406
+ - **Long Tail:** 28,630 words needed for remaining 8.3% 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 | 0.8673 | 0.3719 | N/A | N/A |
432
+ | **mono_64d** | 64 | 0.8240 | 0.2806 | N/A | N/A |
433
+ | **mono_128d** | 128 | 0.6337 | 0.2390 | N/A | N/A |
434
+ | **aligned_32d** | 32 | 0.8673 🏆 | 0.3586 | 0.0220 | 0.1540 |
435
+ | **aligned_64d** | 64 | 0.8240 | 0.2867 | 0.0220 | 0.2300 |
436
+ | **aligned_128d** | 128 | 0.6337 | 0.2384 | 0.0780 | 0.2560 |
437
 
438
  ### Key Findings
439
 
440
+ - **Best Isotropy:** aligned_32d with 0.8673 (more uniform distribution)
441
+ - **Semantic Density:** Average pairwise similarity of 0.2959. Lower values indicate better semantic separation.
442
+ - **Alignment Quality:** Aligned models achieve up to 7.8% 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 | **1.367** | High formulaic/idiomatic content | - |
456
 
457
  ### 6.2 Affix Inventory (Productive Units)
458
 
 
467
 
468
  | Stem | Cohesion | Substitutability | Examples |
469
  |------|----------|------------------|----------|
470
+ | `ther` | 2.76x | 26 contexts | there, other, mother |
471
+ | `tion` | 2.68x | 19 contexts | motion, action, nation |
472
+ | `ount` | 2.74x | 15 contexts | mount, count, counts |
473
+ | `atio` | 2.66x | 15 contexts | ratio, nation, nations |
474
+ | `ctio` | 2.70x | 14 contexts | action, section, actions |
475
+ | `ater` | 2.74x | 11 contexts | later, eater, water |
476
+ | `stat` | 2.72x | 10 contexts | stato, stats, state |
477
+ | `vers` | 2.62x | 11 contexts | verse, covers, rivers |
478
+ | `rati` | 2.70x | 9 contexts | ratio, rating, bharati |
479
+ | `ment` | 2.55x | 9 contexts | cement, ferment, element |
480
+ | `ical` | 2.65x | 8 contexts | typical, medical, optical |
481
+ | `ated` | 2.73x | 7 contexts | dated, stated, related |
482
 
483
  ### 6.4 Affix Compatibility (Co-occurrence)
484
 
 
497
  ### 6.6 Linguistic Interpretation
498
 
499
  > **Automated Insight:**
500
+ The language Bihari languages shows high morphological productivity. The subword models are significantly more efficient than word models, suggesting a rich system of affixation or compounding.
501
+
502
+ > **Note on Idiomaticity:** The high Idiomaticity Gap suggests a large number of frequent multi-word expressions or formulaic sequences that are statistically distinct from their component parts.
503
 
504
  ---
505
  ## 7. Summary & Recommendations
 
511
  | Component | Recommended | Rationale |
512
  |-----------|-------------|-----------|
513
  | Tokenizer | **64k BPE** | Best compression (4.10x) |
514
+ | N-gram | **2-gram** | Lowest perplexity (1,496) |
515
  | Markov | **Context-4** | Highest predictability (96.5%) |
516
  | Embeddings | **100d** | Balanced semantic capture and isotropy |
517
 
 
726
  ---
727
  *Generated by Wikilangs Models Pipeline*
728
 
729
+ *Report Date: 2026-01-03 18:51:04*
models/embeddings/aligned/bh_128d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eea85d4d38b8a1eafd02a0f4039143227a6f20c432fbc6a6fd550310ae30a184
3
+ size 1042277480
models/embeddings/aligned/bh_128d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "bh", "dim": 128, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/bh_128d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fedd4bb9adeb9e4a881c22581a45bc9960409cbbec867b7989c19984436e1845
3
+ size 65664
models/embeddings/aligned/bh_128d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "bh",
3
+ "dimension": 128,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 2989,
7
+ "vocab_size": 17426
8
+ }
models/embeddings/aligned/bh_32d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da9461279c26e26024cae2aab208a18288148ec7cb32f5d81abaebf0820bc365
3
+ size 260894312
models/embeddings/aligned/bh_32d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "bh", "dim": 32, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/bh_32d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8da2bf5ce3276a7611f46c2747071b98c1577b4bc399b70e8744a8ebc0ad2eef
3
+ size 4224
models/embeddings/aligned/bh_32d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "bh",
3
+ "dimension": 32,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 2989,
7
+ "vocab_size": 17426
8
+ }
models/embeddings/aligned/bh_64d.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d06828c33b47aa9adba0339cd59eefcd097881e7258a4a70ad7b1db05455c426
3
+ size 521355368
models/embeddings/aligned/bh_64d.meta.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"lang": "bh", "dim": 64, "max_seq_len": 512, "is_aligned": true}
models/embeddings/aligned/bh_64d.projection.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4482a19a4c700d106e3e055db5374c7ee471b6b83dc203f5be371cc104115ce1
3
+ size 16512
models/embeddings/aligned/bh_64d_metadata.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "language": "bh",
3
+ "dimension": 64,
4
+ "version": "aligned",
5
+ "hub_language": "en",
6
+ "seed_vocab_size": 2989,
7
+ "vocab_size": 17426
8
+ }
models/embeddings/monolingual/bh_128d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b54269438aa2440d60ff4d391e0d7d5fe6f82e878ecccb3b40c6d625a663fdac
3
- size 1042368151
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eea85d4d38b8a1eafd02a0f4039143227a6f20c432fbc6a6fd550310ae30a184
3
+ size 1042277480
models/embeddings/monolingual/bh_128d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 128
13
  },
14
- "vocab_size": 17513
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 128
13
  },
14
+ "vocab_size": 17426
15
  }
models/embeddings/monolingual/bh_32d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8fc27b77183033d6f955d2050d7664babad88301230b75a3837360a073560355
3
- size 260918167
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:da9461279c26e26024cae2aab208a18288148ec7cb32f5d81abaebf0820bc365
3
+ size 260894312
models/embeddings/monolingual/bh_32d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 32
13
  },
14
- "vocab_size": 17513
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 32
13
  },
14
+ "vocab_size": 17426
15
  }
models/embeddings/monolingual/bh_64d.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:3e537a56a3e60b9d66cfebd9427db3b6086b87ba6d254b7115af9f4351f51ee9
3
- size 521401495
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d06828c33b47aa9adba0339cd59eefcd097881e7258a4a70ad7b1db05455c426
3
+ size 521355368
models/embeddings/monolingual/bh_64d_metadata.json CHANGED
@@ -11,5 +11,5 @@
11
  "encoding_method": "rope",
12
  "dim": 64
13
  },
14
- "vocab_size": 17513
15
  }
 
11
  "encoding_method": "rope",
12
  "dim": 64
13
  },
14
+ "vocab_size": 17426
15
  }
models/subword_markov/bh_markov_ctx1_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8a66696e8530ad494e5751234f5b92a3bd0818c8159bfbb9706dcb5b0b509e8c
3
- size 417541
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ac710a309d0b5689e8773f979c3608ce9a0d07db0f59e06b9f80e15194627585
3
+ size 420239
models/subword_markov/bh_markov_ctx1_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_contexts": 4952,
6
- "total_transitions": 4812401
7
  }
 
2
  "context_size": 1,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_contexts": 4950,
6
+ "total_transitions": 4792901
7
  }
models/subword_markov/bh_markov_ctx2_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9c671b706107b700c31f8eb2af638805a0914ba3b55f1ce6cbdde36b234bc3e5
3
- size 1958764
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc8132d17d971115a0b274db540a9f6644dc081f775bd77056355bafb5dd606a
3
+ size 1959380
models/subword_markov/bh_markov_ctx2_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_contexts": 60879,
6
- "total_transitions": 4803696
7
  }
 
2
  "context_size": 2,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_contexts": 60819,
6
+ "total_transitions": 4784188
7
  }
models/subword_markov/bh_markov_ctx3_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d82bb62a4fc528dad712c5f681a9bc4d24b2fb4660975eb2b9fd5687311890b9
3
- size 6336160
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:83798f708304c8896ec57234c8ffe4d4bdd1e596b8997c3d81c249e7e408917d
3
+ size 6309306
models/subword_markov/bh_markov_ctx3_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_contexts": 244880,
6
- "total_transitions": 4794991
7
  }
 
2
  "context_size": 3,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_contexts": 244457,
6
+ "total_transitions": 4775475
7
  }
models/subword_markov/bh_markov_ctx4_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2311c03179b14bfb2afa166ca250ff4c7d2dea7a46e1bf00cd42ce46698deb50
3
- size 15120750
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d804da70fb19cfc45f9559841c58db7cbdbc65b5a134b30fd635d351a7c5938
3
+ size 15126741
models/subword_markov/bh_markov_ctx4_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_contexts": 721221,
6
- "total_transitions": 4786286
7
  }
 
2
  "context_size": 4,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_contexts": 719467,
6
+ "total_transitions": 4766762
7
  }
models/subword_ngram/bh_2gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8ccdb8f666b779178ca817f13ce58935b027cf7910ea8685631ae2bed535b525
3
- size 327904
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:103e2977a1400e02b226091f8d359d50c33a4c1b8f949c0a326a8ff9c80fb2f4
3
+ size 327300
models/subword_ngram/bh_2gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_ngrams": 21796,
6
- "total_ngrams": 4812401
7
  }
 
2
  "n": 2,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_ngrams": 21749,
6
+ "total_ngrams": 4792901
7
  }
models/subword_ngram/bh_3gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4a7163268a6eacd9599d89bf5194ce294c4818e15be4eee3b8f7e92ec8dbbc7d
3
- size 1357005
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:50ed6491fd2b119d9feaecea99e0f03ec4cd34e04e6ec70a95c09e0f54f5f4ca
3
+ size 1358561
models/subword_ngram/bh_3gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_ngrams": 93652,
6
- "total_ngrams": 4803696
7
  }
 
2
  "n": 3,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_ngrams": 93435,
6
+ "total_ngrams": 4784188
7
  }
models/subword_ngram/bh_4gram_subword.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:55ff05f51da64f27b2db568ecf99adc094808b5963aba504d817728e105ceaf2
3
- size 4398449
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb9e46b8c4db1c9131d1b4dbfc2d5c6701f479cf66ef50fe6922288b582d53cd
3
+ size 4384689
models/subword_ngram/bh_4gram_subword_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "bh",
5
- "unique_ngrams": 295453,
6
- "total_ngrams": 4794991
7
  }
 
2
  "n": 4,
3
  "variant": "subword",
4
  "language": "bh",
5
+ "unique_ngrams": 294486,
6
+ "total_ngrams": 4775475
7
  }
models/subword_ngram/bh_5gram_subword.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d26eedca2939b6437bb580f51f3cb3bd6a67e05db9a99a9118ed8107a1444e9
3
+ size 6650702
models/subword_ngram/bh_5gram_subword_metadata.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "n": 5,
3
+ "variant": "subword",
4
+ "language": "bh",
5
+ "unique_ngrams": 421404,
6
+ "total_ngrams": 4766762
7
+ }
models/tokenizer/bh_tokenizer_16k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8e2c6b69efb0696c2593a4e674d531da6cc186970c006b009a5da905fa2c510c
3
- size 609055
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d673d6cc2c6ac90d1d2ca7fc20652e10b454d2918c20da38dd17913cfef448de
3
+ size 609528
models/tokenizer/bh_tokenizer_16k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/bh_tokenizer_32k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a0da439439d6270f2063db9c16c1f281095ac0ed4458b76695c52f7cde9774aa
3
- size 994788
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9e6db347251e4dce2ecc87dd2333e35549ee873caa97fcb62d90ebe18e32f3f6
3
+ size 995212
models/tokenizer/bh_tokenizer_32k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/bh_tokenizer_64k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7e2b6202b97f515f97c09ecfc32ba51664cabebf6f035981a41fc7b7dd5f2814
3
- size 1811363
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f6f2b05c63efaaa5d7198d1e9e86edf6a92a8b6c8a2efb7ecc5f0be6a87caca
3
+ size 1813185
models/tokenizer/bh_tokenizer_64k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/tokenizer/bh_tokenizer_8k.model CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:628a8222f74f579cd406af45e334c28746fbd9325f4a9fcd5c6793ab2534c437
3
- size 420480
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:862301998f448e3ad6016198bd11504ae0eaebc483f07c5422baa0a9e3366e26
3
+ size 420533
models/tokenizer/bh_tokenizer_8k.vocab CHANGED
The diff for this file is too large to render. See raw diff
 
models/vocabulary/bh_vocabulary.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ebd185f91b3ddec379354027ad53da261b000d0e2dccfa4cbfda1efe36a7354d
3
- size 740324
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d376e1ccd56a6b560f8ef35f5dab7b2a85ebb689ec3c060b70d3b927b61d1de5
3
+ size 739098
models/vocabulary/bh_vocabulary_metadata.json CHANGED
@@ -1,17 +1,17 @@
1
  {
2
  "language": "bh",
3
- "vocabulary_size": 38858,
4
  "variant": "full",
5
  "statistics": {
6
- "type_token_ratio": 0.06547537905779276,
7
  "coverage": {
8
- "top_100": 0.41446182158414246,
9
- "top_1000": 0.6701453648682979,
10
- "top_5000": 0.8302279613599608,
11
- "top_10000": 0.8843223122060637
12
  },
13
- "hapax_count": 45677,
14
- "hapax_ratio": 0.5403324066954516,
15
- "total_documents": 8705
16
  }
17
  }
 
1
  {
2
  "language": "bh",
3
+ "vocabulary_size": 38630,
4
  "variant": "full",
5
  "statistics": {
6
+ "type_token_ratio": 0.06557704302490956,
7
  "coverage": {
8
+ "top_100": 0.4151956980527677,
9
+ "top_1000": 0.6710545223497903,
10
+ "top_5000": 0.8308271523523458,
11
+ "top_10000": 0.8847211121182957
12
  },
13
+ "hapax_count": 45795,
14
+ "hapax_ratio": 0.5424341131181523,
15
+ "total_documents": 8713
16
  }
17
  }
models/word_markov/bh_markov_ctx1_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:073f45c70364978f33599e10f7fceb20f9022efb2b123ccffca4352c3c559f2b
3
- size 5186664
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ecde81a1120e9421780f95ce3dc4ab053ecc5e3d4fc1ee6c9aaa63ce280a6469
3
+ size 5178637
models/word_markov/bh_markov_ctx1_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "bh",
5
- "unique_contexts": 84482,
6
- "total_transitions": 1282391
7
  }
 
2
  "context_size": 1,
3
  "variant": "word",
4
  "language": "bh",
5
+ "unique_contexts": 84373,
6
+ "total_transitions": 1278704
7
  }
models/word_markov/bh_markov_ctx2_word.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:a451816cb173e2f9ea644994e76882f796bf7065d5fff1d37c89f78c1ae41b17
3
- size 14774364
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:297824500dbc5072907959bbe396412abce85e0161151273299d228c617723bd
3
+ size 14730272
models/word_markov/bh_markov_ctx2_word_metadata.json CHANGED
@@ -2,6 +2,6 @@
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "bh",
5
- "unique_contexts": 519009,
6
- "total_transitions": 1273686
7
  }
 
2
  "context_size": 2,
3
  "variant": "word",
4
  "language": "bh",
5
+ "unique_contexts": 516874,
6
+ "total_transitions": 1269991
7
  }