Training log

Sansar training runs

Every experiment behind the Sansar models, live and finished, published straight from the machine that trains them.

Scores are bits per byte: how many bits the model needs, on average, to predict each byte of Sanskrit text it never saw in training. Lower is better. How we measure

This is the Sanskrit track of MuseMesh Lab, the log of every model we train: English (7 runs), Math (2 runs), OCR (4 runs).

Live now

Progress

Lower is better on every score chart.

Best ex-Gītā score 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 the milestones listed below. Hollow markers on the top edge are runs scoring above the visible range.

Milestones

Model size and score

Parameters (log scale) against ex-Gītā bits per byte: Sansar runs and other models scored on the same sets. Hover a point for its name.

Tokens per Sanskrit word

How many tokens each tokenizer needs for one word, on average. Fewer is better.

Other models on the same test sets

Scored zero-shot with their own tokenizers and a longer context than ours. Details

External models scored on Sansar's test sets
Model	Parameters	ex-Gītā	clean_v1	Bytes per token
Krutrim-2 12B (instruct)	—	0.512	—	—
Gemma 3 4B pt	3.9B	0.6965	0.7365	6.81
Qwen3-4B-Base	4B	0.7071	0.745	2.78
Qwen3.5-4B-Base	4.2B	0.7114	0.7569	4.66
Llama 3.2 3B	3.2B	0.7122	0.7607	4.68
Sarvam-1	2.5B	0.7465	0.8166	6.12
Qwen3-1.7B-Base	1.7B	0.7872	0.8235	2.78
Llama 3.2 1B	1.2B	0.8072	0.8581	4.68
Qwen3.5-2B-Base	1.9B	0.8117	0.8609	4.66
Gemma 3 1B pt	999.9M	0.8399	0.8778	6.81
Qwen3-0.6B-Base	596M	0.8992	0.9321	2.78

All runs

Published 6 Oct 2026 14:58 UTC · 69 runs

All Sansar (Sanskrit) runs. Column headers sort the table.

RECIPE WD τ0.5+MIX recipe_wd_t05_mix_d125m	6 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 3.51×	7,425	2.2	$2.03	—	—	—
F10 Running f10_slp1_uni8k_d700m_plus_all_v2_3x	5 Oct 2026	d700m 704.1M params	plus_all v2 slice 489.9M words · 3× planned	62,000	25.9	$33.67	—	—	—
RECIPE MIX recipe_mix_d125m	5 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 3.51×	7,425	3.38	—	Discard	0.7236	0.7532
RECIPE ARCH recipe_arch_d125m	5 Oct 2026	d125m 103.4M params	plus_all v2 subset (4 passes) 28.3M words · 4.02×	7,425	3.46	—	Discard	0.7542	0.804
RECIPE R0docs recipe_r0docs_d125m	5 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 4.02×	7,425	3.38	—	Neutral	0.7262	0.7607
RECIPE WD τ0.25 recipe_wd_t025_d125m	4 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 4×	7,425	3.39	—	Keep	0.6946	0.7178
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×	7,425	3.39	—	Winner	0.691	0.7149
RECIPE WD τ1 recipe_wd_t1_d125m	4 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 4×	7,425	3.39	—	Keep	0.6971	0.7225
RECIPE R0 s2 recipe_r0_s2_d125m	4 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 4×	7,425	3.48	—	Keep	0.7136	0.7402
RECIPE R0 recipe_r0_d125m	4 Oct 2026	d125m 97.2M params	plus_all v2 subset (4 passes) 28.3M words · 4×	7,425	3.39	—	Keep	0.7152	0.7427
F10-AB b f10ab_b_d60m	3 Oct 2026	d60m 68.9M params	plus_all band, filtered 90M words · 1.07×	6,358	1.91	—	Neutral	0.6814	—
F10-AB a f10ab_a_d60m	3 Oct 2026	d60m 68.9M params	plus_all band 96M words · 1×	6,358	1.91	—	Neutral	0.6831	—
E-CTX base s2 i9small_base_s2_d60m	3 Oct 2026	d60m 68.9M params	plus_clean subset 1×	6,624	1.99	—	Keep	0.6842	—
E-VOCAB 16k i9small_vocab16k_d60m	3 Oct 2026	d60m 81.2M params	plus_clean subset 1×	5,993	1.95	—	Neutral	0.6843	—
E-VEDIC 3x i9small_vedic3x_d60m	3 Oct 2026	d60m 68.9M params	plus_clean subset + Vedic x3 0.99×	6,635	1.99	—	Keep	0.683	—
E-CTX 1024 i9small_ctx1024_d60m	3 Oct 2026	d60m 68.9M params	plus_clean subset 1×	6,624	2.08	—	Neutral	0.6856	—
F9 f9_slp1_uni8k_d350m_plus_all_3x	3 Oct 2026	d350m 318.4M params	plus_all slice 520.4M words · 3×	103,527	20.53	$26.69	Winner	0.5547	0.5773
E-CTX base i9small_base_d60m	3 Oct 2026	d60m 68.9M params	plus_clean subset 1×	6,624	2.03	—	Keep	0.6825	—
VISION-OCR D s2 visiontest_d_s1338_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,809	0.23	—	Keep	0.7979	—
VISION-OCR D visiontest_d_s1337_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,809	0.23	—	Keep	0.7993	—
VISION-OCR C s2 visiontest_c_s1338_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,847	0.23	—	Neutral	0.8002	—
VISION-OCR B s2 visiontest_b_s1338_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,802	0.23	—	Neutral	0.8004	—
VISION-OCR A s2 visiontest_a_s1338_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,829	0.23	—	Keep	0.7937	—
VISION-OCR C visiontest_c_s1337_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,847	0.23	—	Neutral	0.8015	—
VISION-OCR B visiontest_b_s1337_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,802	0.22	—	Neutral	0.8033	—
VISION-OCR A visiontest_a_s1337_d20m	3 Oct 2026	d20m 27.1M params	plus_clean subset 1×	1,829	0.23	—	Keep	0.7913	—
F8-small f5_slp1_uni8k_d125m_plus_t10	2 Oct 2026	d125m 97.2M params	plus_clean + additions 509.3M words · 0.8×	27,106	12.3	—	Keep	0.6135	—
TOK-v3 v3data tokv3_v3data_d20m	2 Oct 2026	d20m 27.1M params	plus_clean subset 1×	6,689	0.82	—	Discard	0.7191	—
TOK-v3 v3acc tokv3_v3acc_d20m	2 Oct 2026	d20m 27.1M params	plus_clean subset 1×	6,995	0.86	—	Discard	0.7187	—
TOK-v3 v0.1 tokv3_v01_d20m	2 Oct 2026	d20m 27.1M params	plus_clean subset 1×	6,968	0.85	—	Keep	0.7177	—
F7 f7_slp1_uni8k_d350m_plus_clean_2x	2 Oct 2026	d350m 318.4M params	plus_clean slice 448.9M words · 1.84×	54,212	14.01	$18.22	Keep	0.572	0.5937
F6-clean-2x f5_slp1_uni8k_d125m_plus_clean_2x	1 Oct 2026	d125m 97.2M params	plus_clean slice 448.9M words · 1.84×	54,212	24.74	—	Keep	0.6039	0.6254
F6-clean f5_slp1_uni8k_d125m_plus_clean	30 Sep 2026	d125m 97.2M params	plus_clean slice 448.9M words · 0.92×	27,106	12.35	—	Keep	0.6145	—
F4 s2 f4_slp1_uni8k_d125m_muon_arch_s2	30 Sep 2026	d125m 97.2M params	rebuild 5 slice without OCR 446.9M words · 0.93×	27,106	12.38	—	—	0.619	—
F6 f5_slp1_uni8k_d125m_plus	29 Sep 2026	d125m 97.2M params	plus slice	27,106	12.32	—	Discard	0.6142	—
F5-filtered v2 f5_slp1_uni8k_d125m_filtered_v2	27 Sep 2026	d125m 97.2M params	line-filtered slice v2	27,106	12.3	—	Discard	0.6256	—
F5-curated f5_slp1_uni8k_d125m_curated	26 Sep 2026	d125m 97.2M params	curated e-text subset 173M words	27,106	12.34	—	Discard	0.6575	—
F5-filtered f5_slp1_uni8k_d125m_filtered	26 Sep 2026	d125m 97.2M params	line-filtered slice 355M words	27,106	12.28	—	Discard	0.6238	—
F4 f4_slp1_uni8k_d125m_muon_arch	23 Sep 2026	d125m 97.2M params	rebuild 5 slice without OCR 446.9M words · 0.93×	27,106	12.32	—	Keep	0.6189	—
F3-mix12 f3_slp1_uni8k_d125m_mix12	19 Sep 2026	d125m 91.5M params	rebuild 5 slice, 12.5% OCR 505.8M words · 0.81×	27,106	10.89	—	Keep	0.6336	—
F2-noocr f2_slp1_uni8k_d125m_noocr	19 Sep 2026	d125m 91.5M params	rebuild 5 slice without OCR 446.9M words · 0.93×	27,106	10.97	—	Keep	0.6306	—
F2 f2_slp1_uni8k_d125m	18 Sep 2026	d125m 91.5M params	rebuild 5 slice 798.3M words · 0.5×	27,106	10.95	—	Keep	0.6399	—
F1b s2 f1b_slp1_uni8k_d125m_s2	15 Sep 2026	d125m 91.5M params	rebuild 3 slice 414M words · 1×	27,106	10.84	—	Keep	0.641	—
F1b f1b_slp1_uni8k_d125m	14 Sep 2026	d125m 91.5M params	rebuild 3 slice 414M words · 1×	27,106	10.75	—	Keep	0.6372	—
F1 f1_slp1_uni8k_d125m	13 Sep 2026	d125m 91.5M params	rebuild 3 slice 414M words · 1×	27,180	10.78	—	Keep	0.6273	—
F0 f0_slp1_uni8k_d60m	13 Sep 2026	d60m 63.2M params	rebuild 3 slice 414M words · 1×	27,203	7.49	—	Keep	0.6434	—
E15-char-1024 e15_slp1_char_ctx1024	12 Sep 2026	d20m 19.6M params	E8 screening slice 0.59×	19,975	2.12	—	Keep	0.7609	—
E9-16k s2 kg_e9s2_slp1_uni16k	12 Sep 2026	d20m 24M params	E8 screening slice 0.59×	7,151	1.62	Free	Keep	0.766	—
E9-8k s2 kg_e9s2_slp1_uni8k	12 Sep 2026	d20m 22.7M params	E8 screening slice 0.59×	7,891	1.69	Free	Keep	0.7681	—
Gemma 4 E2B CPT Failed kg_gemma4_e2b_cpt30m	12 Sep 2026	— 4B params	Devanagari slice	180	8.07	Free	Discard	2.6118	—
E14-8k e14_slp1_uni8k	12 Sep 2026	d60m 63.2M params	E8 screening slice 0.59×	7,892	2.13	—	Keep	0.7117	—
E14-16k e14_slp1_uni16k	12 Sep 2026	d60m 69.3M params	E8 screening slice 0.59×	7,151	2.1	—	Keep	0.7134	—
E11-bpe-dropout kg_e11_slp1_bpe8k_dropout	12 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	8,306	1.65	Free	Discard	0.814	—
E11-uni-sampling kg_e11_slp1_uni8k_sampling	12 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	12,298	2.03	Free	Discard	0.9526	—
E9-32k e9_slp1_uni32k	12 Sep 2026	d20m 23.1M params	E8 screening slice 0.59×	6,552	0.76	—	Discard	0.7671	—
E9-16k e9_slp1_uni16k	12 Sep 2026	d20m 24M params	E8 screening slice 0.59×	7,151	0.91	—	Keep	0.7556	—
E10-bpe16k kg_slp1_bpe16k	12 Sep 2026	d20m 27.3M params	E8 screening slice 0.59×	6,441	1.28	Free	Discard	0.7762	—
E9-8k e9_slp1_uni8k	12 Sep 2026	d20m 22.7M params	E8 screening slice 0.59×	7,892	0.95	—	Keep	0.7552	—
E10-uni16k kg_slp1_uni16k_d20m	12 Sep 2026	d20m 27.3M params	E8 screening slice 0.59×	7,151	1.41	Free	Keep	0.7626	—
E9-4k e9_slp1_uni4k	12 Sep 2026	d20m 23M params	E8 screening slice 0.59×	8,746	1.07	—	Discard	0.7581	—
E8-slp1-uni e8_slp1_uni8k	11 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	7,948	0.88	—	Keep	0.7537	—
E8-slp1-char e8_slp1_char	11 Sep 2026	d20m 19.4M params	E8 screening slice 0.59×	19,995	1.85	—	Keep	0.7503	—
E8-slp1-bpe e8_slp1_bpe8k	11 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	7,083	0.78	—	Discard	0.7707	—
E8-deva-uni e8_deva_uni8k_syms	11 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	7,635	0.84	—	Winner	0.7505	—
E8-deva-bpe e8_deva_bpe8k_nosyms	11 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	6,854	0.77	—	Discard	0.7713	—
E8-deva-bpe-sym e8_deva_bpe8k_syms	11 Sep 2026	d20m 23.2M params	E8 screening slice 0.59×	7,121	0.78	—	Discard	0.7719	—
Smoke test smoke_d20m	11 Sep 2026	d20m 23.2M params	early test slice	8,074	1	—	—	—	—
RECIPE WD τ0.5 s2 Queued recipe_wd_t05_s2_d125m	—	—	plus_all v2 subset (4 passes) 28.3M words	0	—	—	—	—	—
RECIPE WD τ0.5 muon-only Queued recipe_wd_t05_muononly_d125m	—	—	plus_all v2 subset (4 passes) 28.3M words	0	—	—	—	—	—

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