Sansar: Sanskrit language models - a MuseMesh Collection

MuseMesh 's Collections

Mume: English and Math language models

Sansar: Sanskrit language models

Sansar: Sanskrit language models

updated 13 days ago

Sanskrit-only models trained from scratch, their tokenizer and the open Sanskrit corpus.

MuseMesh/sansar-700m

Text Generation • 0.7B • Updated 9 days ago • 341 • 1
MuseMesh/sansar-350m

Text Generation • 0.3B • Updated 15 days ago • 514 • 1

Note Best model: ex-Gita 0.5547 bits per byte
MuseMesh/sansar-125m

Text Generation • 97.2M • Updated 15 days ago • 495 • 1
MuseMesh/sansar-60m

Text Generation • 63.2M • Updated 15 days ago • 495 • 1
MuseMesh/sansar-20m

Text Generation • 27.1M • Updated 15 days ago • 486 • 1
MuseMesh/sansar-sanskrit-corpus

Viewer • Updated 18 days ago • 3.66M • 104 • 1

Note 252M words, one config per licence
MuseMesh/sansar-sanskrit-tokenizer

Updated 21 days ago • 1

Note 8k SLP1 unigram + Devanagari wrapper
MuseMesh/sansar-ocr-2b

Image-Text-to-Text • 2B • Updated 13 days ago • 40 • 1

Note Printed Sanskrit page OCR (Qwen3.5-2B fine-tune); 1.05% median letter error vs Google Vision 1.69%. Trained only on clean e-texts, born-digital text layers and synthetic pages.