Training machine offline · last report 17 days ago

Sanskrit training runs

Sansar: language models for Sanskrit only, trained from scratch on an open 252M-word corpus with a Sanskrit tokenizer of their own.

Headline score: ex-Gītā · lower is better Bits per byte pooled over four held-out Sanskrit sets: classical prose and verse, web text and the Ṛgveda. The Bhagavad-gītā is left out because the models partly memorise it.Bits per byte: how many bits the model needs, on average, to predict each byte of text it never saw in training. Lower is better, and it compares models with different tokenizers fairly, because every model is charged for the same bytes. · Try the demo · how Sanskrit is measured

Every score on this track
ex-Gītā (lower is better)
Bits per byte pooled over four held-out Sanskrit sets: classical prose and verse, web text and the Ṛgveda. The Bhagavad-gītā is left out because the models partly memorise it.
ex-Gītā, clean_v1 (lower is better)
The same four sets minus every item that could overlap the training text. Stricter; compare only with other clean_v1 numbers.
Pooled, five sets (lower is better)
All five held-out sets pooled, the Gītā included.
Validation split (lower is better)
A slice of the training corpus held back from training, scored during the run.

Progress

Lower is better on the headline.

Best ex-Gītā over time

Each dot is a scored run, placed at the day it ended; the line is the best score so far. Numbered lines are this track's milestones, listed below.

Milestones
  1. 12 Oct 2026 · Sansar 700M v0.2.0 (F11) on Hugging Face
  2. 11 Oct 2026 · F11: best model so far (ex-Gītā 0.5262, 0.8% better than F10)
  3. 8 Oct 2026 · F11 (F10 with four passes) starts training on AWS
  4. 7 Oct 2026 · The lab moves to its own home: saansar.com
  5. 7 Oct 2026 · Sansar 700M public on Hugging Face (F10)
  6. 7 Oct 2026 · F10: first 700M model, best so far (ex-Gītā 0.5307, 4.3% better than F9)
  7. 6 Oct 2026 · Sansar 350m compared with open base models up to 4B
  8. 5 Oct 2026 · F10 (700M) starts training
  9. 4 Oct 2026 · A 252M-word Sanskrit corpus public on Hugging Face
  10. 4 Oct 2026 · Sansar 20M, 60M, 125M and 350M public on Hugging Face
  11. 4 Oct 2026 · F9: best model so far (ex-Gītā 0.5547, 3.0% better than F7)
  12. 3 Oct 2026 · F7: first 350M model, first cloud run, best so far (ex-Gītā 0.5720, 5.3% better than F6-clean-2x)
  13. 2 Oct 2026 · F6-clean-2x: best model so far (ex-Gītā 0.6039, 1.7% better than F6)
  14. 1 Oct 2026 · First release: the Sansar tokenizer on Hugging Face
  15. 24 Sep 2026 · F4: the Muon + architecture recipe transfers to Sanskrit
  16. 20 Sep 2026 · OCR text hurts in proportion to its share
  17. 14 Sep 2026 · Held-out masking: the honest baseline F1b
  18. 14 Sep 2026 · F1: first 125M model, best so far (ex-Gītā 0.6273, 2.5% better than F0)
  19. 13 Sep 2026 · F0: first model on the full corpus (60M)
  20. 12 Sep 2026 · Tokenizer frozen: SLP1 unigram 8k
  21. 12 Sep 2026 · E14-16k: first 60M model, best so far (ex-Gītā 0.7134, 4.9% better than E8-slp1-char)
  22. 12 Sep 2026 · E8-deva-uni: best model so far (ex-Gītā 0.7505, 2.7% better than E8-deva-bpe)
  23. 11 Sep 2026 · First end-to-end training run on the home GPU

Model size and ex-Gītā

Parameters (log scale) against ex-Gītā in bits per byte: this track's runs and other models measured the same way. Hover a point for its name.

Reference models

Other models measured on the same sets, for comparison. Notes say how they were measured.

Reference models scored on the Sanskrit track's test sets
ModelParametersex-Gītāex-Gītā, clean_v1
Krutrim-2 12B (instruct) Scored in an earlier pass (4-bit weights, verse references stripped), so not strictly comparable; it beats Sansar 350m there.—0.512—
Gemma 3 4B pt3.9B0.69650.7365
Qwen3-4B-Base4B0.70710.745
Qwen3.5-4B-Base4.2B0.71140.7569
Llama 3.2 3B3.2B0.71220.7607
Sarvam-12.5B0.74650.8166
Qwen3-1.7B-Base1.7B0.78720.8235
Llama 3.2 1B1.2B0.80720.8581
Qwen3.5-2B-Base1.9B0.81170.8609
Gemma 3 1B pt999.9M0.83990.8778
Qwen3-0.6B-Base596M0.89920.9321

Sanskrit runs

Published · 83 runs

All Sanskrit runs. Column headers sort the table.
Data
F11
f11_slp1_uni8k_d700m_plus_all_v2_4x
8 Oct 2026 d700m
704.1M params
plus_all v2 slice
489.9M words · 4×
71.48 $133.97 Winner 0.5262 0.5507 0.5094 0.4389
F10
f10_slp1_uni8k_d700m_plus_all_v2_3x
5 Oct 2026 d700m
704.1M params
plus_all v2 slice
489.9M words · 3×
40.45 $52.59 Winner 0.5307 0.5541 0.514 0.4486
RECIPE WD τ0.5
recipe_wd_t05_d125m
4 Oct 2026 d125m
97.2M params
plus_all v2 subset (4 passes)
28.3M words · 4×
3.39 — Winner 0.691 0.7149 0.6822 0.6741
F9
f9_slp1_uni8k_d350m_plus_all_3x
3 Oct 2026 d350m
318.4M params
plus_all slice
520.4M words · 3×
20.53 $26.69 Winner 0.5547 0.5773 0.5394 0.5033

Cost “—” means no cloud bill (our own GPU at home) or a cost that was not recorded; “Free” is free cloud compute. GPU-h is GPU hours.