Training log

Training machine offline · last report 17 days ago

Sanskrit-first language models, built in the open

We train language models from scratch, starting with Sanskrit, and publish every run live from the machines that train them: what we tested, what it scored, what it cost, and what failed.

Tracks

Tracks

Published

Sanskrit

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

Best · ex-Gītā
0.5262 bits/byte
F11 · lower is better
Runs
83
Last activity
3 Nov 2026

English

125M-class English models on FineWeb web text. The proving ground: optimisers and architecture are tested here first, against a known GPT-2 reference.

Best · FineWeb validation
1.10625 bits/byte
EN1-full · lower is better
Runs
7
Last activity
24 Sep 2026

Math

A 125M math model trained on mathematical web pages with the English recipe, and what fine-tuning on worked solutions adds.

Best · OpenWebMath held-out, clean
1.06595 bits/byte
MATH0 · lower is better
Runs
5
Last activity
7 Oct 2026

OCR

Reading scanned Sanskrit books: an open vision-language model fine-tuned for page OCR, and the bulk OCR jobs that turn scans into corpus text.

Best · Median letter error rate
0.0096
OCR-VLM v1 · lower is better
Runs
5
Last activity
8 Oct 2026

Voice

Sanskrit text-to-speech. The goal is two voices of our own: one that chants verse by its metre, one that reads prose. Daily Sanskrit already chants with Vāgdhenu by Prathosh A P (IISc), credited there. No audio from these experiments is published.

Best · Letter error, bench
0.007
TTS-E3 · lower is better
Runs
8
Last activity
10 Oct 2026

Today's Sanskrit

Something from the corpus every day, with translation, meaning and pronunciation

Mantra सं गच्छध्वं सं वदध्वं सं वो मनांसि जानताम् । Saṃ gacchadhvam: the last hymn of the Ṛgveda Come together, speak together, let your minds be of one accord, just as the gods of old, in agreement, sat near to receive their share. Read it, word by word

Timeline

Each dot is a run, placed at the day it ended (or started, while it runs).

Runs by track over time

One row per track. Hover a dot for the run and its headline score; numbered lines are the milestones listed below.

Milestones
  1. 12 Oct 2026 · Sanskrit Sansar 700M v0.2.0 (F11) on Hugging Face
  2. 11 Oct 2026 · Sanskrit F11: best model so far (ex-Gītā 0.5262, 0.8% better than F10)
  3. 9 Oct 2026 · Voice TTS-E1: the first Sanskrit text-to-speech fine-tune starts on a cloud GPU
  4. 9 Oct 2026 · OCR Sansar OCR 2B public on Hugging Face (OCR-VLM v1.2 clean)
  5. 8 Oct 2026 · Voice Voice bake-off: open Sanskrit text-to-speech models scored for letter error and vowel length
  6. 8 Oct 2026 · Sanskrit F11 (F10 with four passes) starts training on AWS
  7. 7 Oct 2026 · Math MATH0-MID: arithmetic drills in pre-training lift DeepMind maths to 19.5%; GSM8K unchanged
  8. 7 Oct 2026 · Math MATH1-SFT: 35x the worked solutions, no arithmetic gain; DeepMind maths exact match measured
  9. 7 Oct 2026 · Sanskrit The lab moves to its own home: saansar.com
  10. 7 Oct 2026 · Sanskrit Sansar 700M public on Hugging Face (F10)
  11. 7 Oct 2026 · Sanskrit F10: first 700M model, best so far (ex-Gītā 0.5307, 4.3% better than F9)
  12. 6 Oct 2026 · OCR OCR-VLM v1 beats Google Vision on Sanskrit pages
  13. 6 Oct 2026 · Math Mume Math 125M public on Hugging Face (MATH0 and MATH0-SFT)
  14. 6 Oct 2026 · English Mume English 125M public on Hugging Face (EN0 and EN1-full)
  15. 6 Oct 2026 · Sanskrit Sansar 350m compared with open base models up to 4B
  16. 5 Oct 2026 · OCR OCR-VLM pilot: first OCR run
  17. 5 Oct 2026 · Sanskrit F10 (700M) starts training
  18. 4 Oct 2026 · Sanskrit A 252M-word Sanskrit corpus public on Hugging Face
  19. 4 Oct 2026 · Sanskrit Sansar 20M, 60M, 125M and 350M public on Hugging Face
  20. 4 Oct 2026 · Sanskrit F9: best model so far (ex-Gītā 0.5547, 3.0% better than F7)
  21. 3 Oct 2026 · Sanskrit F7: first 350M model, first cloud run, best so far (ex-Gītā 0.5720, 5.3% better than F6-clean-2x)
  22. 2 Oct 2026 · Sanskrit F6-clean-2x: best model so far (ex-Gītā 0.6039, 1.7% better than F6)
  23. 1 Oct 2026 · Math MATH0-SFT: fine-tuned on worked solutions, GSM8K 2.43%
  24. 1 Oct 2026 · Sanskrit First release: the Sansar tokenizer on Hugging Face
  25. 29 Sep 2026 · Math MATH0: first math model, 11% to 63% better than English on math
  26. 24 Sep 2026 · English EN1-full beats a plain AdamW GPT-2 124M at equal tokens
  27. 24 Sep 2026 · Sanskrit F4: the Muon + architecture recipe transfers to Sanskrit
  28. 23 Sep 2026 · English EN1: Muon and a modern block stack, 4.8% better in English
  29. 22 Sep 2026 · English EN0: the pipeline lands on the GPT-2 124M curve in English
  30. 20 Sep 2026 · Sanskrit OCR text hurts in proportion to its share
  31. 14 Sep 2026 · Sanskrit Held-out masking: the honest baseline F1b
  32. 14 Sep 2026 · Sanskrit F1: first 125M model, best so far (ex-Gītā 0.6273, 2.5% better than F0)
  33. 13 Sep 2026 · Sanskrit F0: first model on the full corpus (60M)
  34. 12 Sep 2026 · Sanskrit Tokenizer frozen: SLP1 unigram 8k
  35. 12 Sep 2026 · Sanskrit E14-16k: first 60M model, best so far (ex-Gītā 0.7134, 4.9% better than E8-slp1-char)
  36. 12 Sep 2026 · Sanskrit E8-deva-uni: best model so far (ex-Gītā 0.7505, 2.7% better than E8-deva-bpe)
  37. 11 Sep 2026 · Sanskrit First end-to-end training run on the home GPU

Runs

108 runs across 5 tracks

Every published training run. Column headers sort the table.
OCR-VLM v1.2 clean
ocrvlm_v12_clean_q35
OCR 8 Oct 2026 — 0.0106
Median letter error rate
Winner 7.59 $9.86
F11
f11_slp1_uni8k_d700m_plus_all_v2_4x
Sanskrit 8 Oct 2026 d700m
704.1M params
0.5262
ex-Gītā, bits/byte
Winner 71.48 $133.97
F10
f10_slp1_uni8k_d700m_plus_all_v2_3x
Sanskrit 5 Oct 2026 d700m
704.1M params
0.5307
ex-Gītā, bits/byte
Winner 40.45 $52.59
OCR-VLM v1
ocrvlm_v1_q35
OCR 5 Oct 2026 — 0.0096
Median letter error rate
Winner 7.33 $9.53
RECIPE WD τ0.5
recipe_wd_t05_d125m
Sanskrit 4 Oct 2026 d125m
97.2M params
0.691
ex-Gītā, bits/byte
Winner 3.39 —
F9
f9_slp1_uni8k_d350m_plus_all_3x
Sanskrit 3 Oct 2026 d350m
318.4M params
0.5547
ex-Gītā, bits/byte
Winner 20.53 $26.69
MATH0
math0_shared32k_d125m_ctx1024
Math 28 Sep 2026 d125m
134.1M params
1.06595
OpenWebMath held-out, clean, bits/byte
Winner 14.82 —
EN1-full
en1_full_d125m_ctx1024
English 23 Sep 2026 d125m
134.1M params
1.10625
FineWeb validation, bits/byte
Winner 14.85 —

Headline scores differ by track (each track page explains its own); compare them within a track. 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.