Training log

MuseMesh Lab: every training run, live

Every model MuseMesh trains from scratch, and every experiment behind them, published straight from the machines that train them: what we tested, what we learned, and what it cost.

Tracks

Published

Sanskrit

1 live

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
0.5547 bits/byte
F9 · lower is better
Runs
69
Last activity
6 Oct 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
2
Last activity
1 Oct 2026

OCR

1 live

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
4
Last activity
6 Oct 2026

Live now

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. 11 Sep 2026 · Sanskrit First end-to-end training run on the home GPU
  2. 12 Sep 2026 · Sanskrit Tokenizer frozen: SLP1 unigram 8k
  3. 13 Sep 2026 · Sanskrit F0: first model on the full corpus (60M)
  4. 14 Sep 2026 · Sanskrit Held-out masking: the honest baseline F1b
  5. 20 Sep 2026 · Sanskrit OCR text hurts in proportion to its share
  6. 22 Sep 2026 · English EN0: the pipeline lands on the GPT-2 124M curve in English
  7. 23 Sep 2026 · English EN1: Muon and a modern block stack, 4.8% better in English
  8. 24 Sep 2026 · English EN1-full beats a plain AdamW GPT-2 124M at equal tokens
  9. 24 Sep 2026 · Sanskrit F4: the Muon + architecture recipe transfers to Sanskrit
  10. 29 Sep 2026 · Math MATH0: first math model, 11% to 63% better than English on math
  11. 1 Oct 2026 · Math MATH0-SFT: fine-tuned on worked solutions, GSM8K 2.43%
  12. 1 Oct 2026 · Sanskrit First release: the Sansar tokenizer on Hugging Face
  13. 3 Oct 2026 · Sanskrit F7: first 350M model, first cloud run
  14. 4 Oct 2026 · Sanskrit F9: best model so far (ex-Gita 0.5547)
  15. 4 Oct 2026 · Sanskrit Models (20M to 350M) and a 252M-word corpus public on Hugging Face
  16. 5 Oct 2026 · Sanskrit F10 (700M) starts training
  17. 6 Oct 2026 · OCR OCR-VLM v1 beats Google Vision on Sanskrit pages
  18. 6 Oct 2026 · Sanskrit Sansar 350m compared with open base models up to 4B

All runs

82 runs across 4 tracks

Every published training run. Column headers sort the table.
OCR pilot test
ocr_pilot100k_test
OCR 6 Oct 2026 — — — 0.57 $0.52
RECIPE WD τ0.5+MIX
recipe_wd_t05_mix_d125m
Sanskrit 6 Oct 2026 d125m
97.2M params
— — 2.2 $2.03
F10 Running
f10_slp1_uni8k_d700m_plus_all_v2_3x
Sanskrit 5 Oct 2026 d700m
704.1M params
— — 25.9 $33.67
RECIPE MIX
recipe_mix_d125m
Sanskrit 5 Oct 2026 d125m
97.2M params
0.7236
ex-Gītā, bits/byte
Discard 3.38 —
RECIPE ARCH
recipe_arch_d125m
Sanskrit 5 Oct 2026 d125m
103.4M params
0.7542
ex-Gītā, bits/byte
Discard 3.46 —
OCR-VLM v1
ocrvlm_v1_q35
OCR 5 Oct 2026 — 0.0096
Median letter error rate
Winner 7.33 $9.53
RECIPE R0docs
recipe_r0docs_d125m
Sanskrit 5 Oct 2026 d125m
97.2M params
0.7262
ex-Gītā, bits/byte
Neutral 3.38 —
RECIPE WD τ0.25
recipe_wd_t025_d125m
Sanskrit 4 Oct 2026 d125m
97.2M params
0.6946
ex-Gītā, bits/byte
Keep 3.39 —
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 —
OCR-VLM pilot
ocrvlm_pilot_q35
OCR 4 Oct 2026 — — Keep 1.37 $1.80
RECIPE WD τ1
recipe_wd_t1_d125m
Sanskrit 4 Oct 2026 d125m
97.2M params
0.6971
ex-Gītā, bits/byte
Keep 3.39 —
RECIPE R0 s2
recipe_r0_s2_d125m
Sanskrit 4 Oct 2026 d125m
97.2M params
0.7136
ex-Gītā, bits/byte
Keep 3.48 —
RECIPE R0
recipe_r0_d125m
Sanskrit 4 Oct 2026 d125m
97.2M params
0.7152
ex-Gītā, bits/byte
Keep 3.39 —
F10-AB b
f10ab_b_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.6814
ex-Gītā, bits/byte
Neutral 1.91 —
F10-AB a
f10ab_a_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.6831
ex-Gītā, bits/byte
Neutral 1.91 —
E-CTX base s2
i9small_base_s2_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.6842
ex-Gītā, bits/byte
Keep 1.99 —
E-VOCAB 16k
i9small_vocab16k_d60m
Sanskrit 3 Oct 2026 d60m
81.2M params
0.6843
ex-Gītā, bits/byte
Neutral 1.95 —
E-VEDIC 3x
i9small_vedic3x_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.683
ex-Gītā, bits/byte
Keep 1.99 —
E-CTX 1024
i9small_ctx1024_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.6856
ex-Gītā, bits/byte
Neutral 2.08 —
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
E-CTX base
i9small_base_d60m
Sanskrit 3 Oct 2026 d60m
68.9M params
0.6825
ex-Gītā, bits/byte
Keep 2.03 —
VISION-OCR D s2
visiontest_d_s1338_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.7979
ex-Gītā, bits/byte
Keep 0.23 —
VISION-OCR D
visiontest_d_s1337_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.7993
ex-Gītā, bits/byte
Keep 0.23 —
VISION-OCR C s2
visiontest_c_s1338_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.8002
ex-Gītā, bits/byte
Neutral 0.23 —
VISION-OCR B s2
visiontest_b_s1338_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.8004
ex-Gītā, bits/byte
Neutral 0.23 —
VISION-OCR A s2
visiontest_a_s1338_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.7937
ex-Gītā, bits/byte
Keep 0.23 —
VISION-OCR C
visiontest_c_s1337_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.8015
ex-Gītā, bits/byte
Neutral 0.23 —
VISION-OCR B
visiontest_b_s1337_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.8033
ex-Gītā, bits/byte
Neutral 0.22 —
VISION-OCR A
visiontest_a_s1337_d20m
Sanskrit 3 Oct 2026 d20m
27.1M params
0.7913
ex-Gītā, bits/byte
Keep 0.23 —
F8-small
f5_slp1_uni8k_d125m_plus_t10
Sanskrit 2 Oct 2026 d125m
97.2M params
0.6135
ex-Gītā, bits/byte
Keep 12.3 —
TOK-v3 v3data
tokv3_v3data_d20m
Sanskrit 2 Oct 2026 d20m
27.1M params
0.7191
ex-Gītā, bits/byte
Discard 0.82 —
TOK-v3 v3acc
tokv3_v3acc_d20m
Sanskrit 2 Oct 2026 d20m
27.1M params
0.7187
ex-Gītā, bits/byte
Discard 0.86 —
TOK-v3 v0.1
tokv3_v01_d20m
Sanskrit 2 Oct 2026 d20m
27.1M params
0.7177
ex-Gītā, bits/byte
Keep 0.85 —
F7
f7_slp1_uni8k_d350m_plus_clean_2x
Sanskrit 2 Oct 2026 d350m
318.4M params
0.572
ex-Gītā, bits/byte
Keep 14.01 $18.22
F6-clean-2x
f5_slp1_uni8k_d125m_plus_clean_2x
Sanskrit 1 Oct 2026 d125m
97.2M params
0.6039
ex-Gītā, bits/byte
Keep 24.74 —
MATH0-SFT
math0_sft
Math 1 Oct 2026 d125m
134.1M params
2.43%
GSM8K exact match
Keep 0.22 —
F6-clean
f5_slp1_uni8k_d125m_plus_clean
Sanskrit 30 Sep 2026 d125m
97.2M params
0.6145
ex-Gītā, bits/byte
Keep 12.35 —
F4 s2
f4_slp1_uni8k_d125m_muon_arch_s2
Sanskrit 30 Sep 2026 d125m
97.2M params
0.619
ex-Gītā, bits/byte
— 12.38 —
F6
f5_slp1_uni8k_d125m_plus
Sanskrit 29 Sep 2026 d125m
97.2M params
0.6142
ex-Gītā, bits/byte
Discard 12.32 —
MATH0
math0_shared32k_d125m_ctx1024
Math 28 Sep 2026 d125m
134.1M params
1.06595
OpenWebMath held-out, clean, bits/byte
Winner 14.82 —
F5-filtered v2
f5_slp1_uni8k_d125m_filtered_v2
Sanskrit 27 Sep 2026 d125m
97.2M params
0.6256
ex-Gītā, bits/byte
Discard 12.3 —
F5-curated
f5_slp1_uni8k_d125m_curated
Sanskrit 26 Sep 2026 d125m
97.2M params
0.6575
ex-Gītā, bits/byte
Discard 12.34 —
F5-filtered
f5_slp1_uni8k_d125m_filtered
Sanskrit 26 Sep 2026 d125m
97.2M params
0.6238
ex-Gītā, bits/byte
Discard 12.28 —
EN1-full
en1_full_d125m_ctx1024
English 23 Sep 2026 d125m
134.1M params
1.10625
FineWeb validation, bits/byte
Winner 14.85 —
F4
f4_slp1_uni8k_d125m_muon_arch
Sanskrit 23 Sep 2026 d125m
97.2M params
0.6189
ex-Gītā, bits/byte
Keep 12.32 —
EN1-muon_arch_wd
en1_muon_arch_wd_d125m_ctx1024
English 23 Sep 2026 d125m
158.7M params
1.1909
FineWeb validation, bits/byte
Discard 5.54 —
EN1-muon_arch
en1_muon_arch_d125m_ctx1024
English 22 Sep 2026 d125m
134.1M params
1.16018
FineWeb validation, bits/byte
Keep 5.44 —
EN1-arch
en1_arch_d125m_ctx1024
English 22 Sep 2026 d125m
134.1M params
1.19028
FineWeb validation, bits/byte
Keep 5.16 —
EN1-ctrl
en1_adamw_ctrl_d125m_ctx1024
English 22 Sep 2026 d125m
110.3M params
1.21874
FineWeb validation, bits/byte
Keep 4.95 —
EN1-muon
en1_muon_d125m_ctx1024
English 22 Sep 2026 d125m
110.3M params
1.19873
FineWeb validation, bits/byte
Keep 5.33 —
EN0
en0_shared32k_d125m_ctx1024
English 21 Sep 2026 d125m
110.3M params
1.15174
FineWeb validation, bits/byte
Keep 13.89 —
F3-mix12
f3_slp1_uni8k_d125m_mix12
Sanskrit 19 Sep 2026 d125m
91.5M params
0.6336
ex-Gītā, bits/byte
Keep 10.89 —
F2-noocr
f2_slp1_uni8k_d125m_noocr
Sanskrit 19 Sep 2026 d125m
91.5M params
0.6306
ex-Gītā, bits/byte
Keep 10.97 —
F2
f2_slp1_uni8k_d125m
Sanskrit 18 Sep 2026 d125m
91.5M params
0.6399
ex-Gītā, bits/byte
Keep 10.95 —
F1b s2
f1b_slp1_uni8k_d125m_s2
Sanskrit 15 Sep 2026 d125m
91.5M params
0.641
ex-Gītā, bits/byte
Keep 10.84 —
F1b
f1b_slp1_uni8k_d125m
Sanskrit 14 Sep 2026 d125m
91.5M params
0.6372
ex-Gītā, bits/byte
Keep 10.75 —
F1
f1_slp1_uni8k_d125m
Sanskrit 13 Sep 2026 d125m
91.5M params
0.6273
ex-Gītā, bits/byte
Keep 10.78 —
F0
f0_slp1_uni8k_d60m
Sanskrit 13 Sep 2026 d60m
63.2M params
0.6434
ex-Gītā, bits/byte
Keep 7.49 —
E15-char-1024
e15_slp1_char_ctx1024
Sanskrit 12 Sep 2026 d20m
19.6M params
0.7609
ex-Gītā, bits/byte
Keep 2.12 —
E9-16k s2
kg_e9s2_slp1_uni16k
Sanskrit 12 Sep 2026 d20m
24M params
0.766
ex-Gītā, bits/byte
Keep 1.62 Free
E9-8k s2
kg_e9s2_slp1_uni8k
Sanskrit 12 Sep 2026 d20m
22.7M params
0.7681
ex-Gītā, bits/byte
Keep 1.69 Free
Gemma 4 E2B CPT Failed
kg_gemma4_e2b_cpt30m
Sanskrit 12 Sep 2026 —
4B params
2.6118
ex-Gītā, bits/byte
Discard 8.07 Free
E14-8k
e14_slp1_uni8k
Sanskrit 12 Sep 2026 d60m
63.2M params
0.7117
ex-Gītā, bits/byte
Keep 2.13 —
E14-16k
e14_slp1_uni16k
Sanskrit 12 Sep 2026 d60m
69.3M params
0.7134
ex-Gītā, bits/byte
Keep 2.1 —
E11-bpe-dropout
kg_e11_slp1_bpe8k_dropout
Sanskrit 12 Sep 2026 d20m
23.2M params
0.814
ex-Gītā, bits/byte
Discard 1.65 Free
E11-uni-sampling
kg_e11_slp1_uni8k_sampling
Sanskrit 12 Sep 2026 d20m
23.2M params
0.9526
ex-Gītā, bits/byte
Discard 2.03 Free
E9-32k
e9_slp1_uni32k
Sanskrit 12 Sep 2026 d20m
23.1M params
0.7671
ex-Gītā, bits/byte
Discard 0.76 —
E9-16k
e9_slp1_uni16k
Sanskrit 12 Sep 2026 d20m
24M params
0.7556
ex-Gītā, bits/byte
Keep 0.91 —
E10-bpe16k
kg_slp1_bpe16k
Sanskrit 12 Sep 2026 d20m
27.3M params
0.7762
ex-Gītā, bits/byte
Discard 1.28 Free
E9-8k
e9_slp1_uni8k
Sanskrit 12 Sep 2026 d20m
22.7M params
0.7552
ex-Gītā, bits/byte
Keep 0.95 —
E10-uni16k
kg_slp1_uni16k_d20m
Sanskrit 12 Sep 2026 d20m
27.3M params
0.7626
ex-Gītā, bits/byte
Keep 1.41 Free
E9-4k
e9_slp1_uni4k
Sanskrit 12 Sep 2026 d20m
23M params
0.7581
ex-Gītā, bits/byte
Discard 1.07 —
E8-slp1-uni
e8_slp1_uni8k
Sanskrit 11 Sep 2026 d20m
23.2M params
0.7537
ex-Gītā, bits/byte
Keep 0.88 —
E8-slp1-char
e8_slp1_char
Sanskrit 11 Sep 2026 d20m
19.4M params
0.7503
ex-Gītā, bits/byte
Keep 1.85 —
E8-slp1-bpe
e8_slp1_bpe8k
Sanskrit 11 Sep 2026 d20m
23.2M params
0.7707
ex-Gītā, bits/byte
Discard 0.78 —
E8-deva-uni
e8_deva_uni8k_syms
Sanskrit 11 Sep 2026 d20m
23.2M params
0.7505
ex-Gītā, bits/byte
Winner 0.84 —
E8-deva-bpe
e8_deva_bpe8k_nosyms
Sanskrit 11 Sep 2026 d20m
23.2M params
0.7713
ex-Gītā, bits/byte
Discard 0.77 —
E8-deva-bpe-sym
e8_deva_bpe8k_syms
Sanskrit 11 Sep 2026 d20m
23.2M params
0.7719
ex-Gītā, bits/byte
Discard 0.78 —
Smoke test
smoke_d20m
Sanskrit 11 Sep 2026 d20m
23.2M params
— — 1 —
RECIPE WD τ0.5 s2 Queued
recipe_wd_t05_s2_d125m
Sanskrit — — — — — —
RECIPE WD τ0.5 muon-only Queued
recipe_wd_t05_muononly_d125m
Sanskrit — — — — — —
OCR pilot 100k Running
ocr_pilot100k
OCR — — — — — $5.81

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.