Create tokenizer/README.md
Browse files- tokenizer/README.md +19 -0
tokenizer/README.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Sanchari Tokenizer
|
| 2 |
+
|
| 3 |
+
This folder contains scripts and placeholder artifacts for the Sanchari tokenizer.
|
| 4 |
+
|
| 5 |
+
The tokenizer is based on SentencePiece (Unigram or BPE) with a ~50k vocabulary optimized for:
|
| 6 |
+
- English (India)
|
| 7 |
+
- Hindi
|
| 8 |
+
- Telugu
|
| 9 |
+
- Mixed-script content
|
| 10 |
+
- Code + instruction-level text
|
| 11 |
+
|
| 12 |
+
Tokenization goals:
|
| 13 |
+
- Normalize Unicode (NFKC)
|
| 14 |
+
- Efficient segmentation for Indic languages
|
| 15 |
+
- Stable handling of whitespace, punctuation, emojis, and mixed-language text
|
| 16 |
+
|
| 17 |
+
Final tokenizer files (`sanchari_spm.model` and `sanchari_spm.vocab`) will be generated after dataset aggregation.
|
| 18 |
+
|
| 19 |
+
This version contains **placeholders only** for investor preview.
|