Training log

Training machine offline · last report 17 days ago

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

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
    ModelParametersex-Gītāclean_v1Bytes per token
    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 pt3.9B0.69650.73656.81
    Qwen3-4B-Base4B0.70710.7452.78
    Qwen3.5-4B-Base4.2B0.71140.75694.66
    Llama 3.2 3B3.2B0.71220.76074.68
    Sarvam-12.5B0.74650.81666.12
    Qwen3-1.7B-Base1.7B0.78720.82352.78
    Llama 3.2 1B1.2B0.80720.85814.68
    Qwen3.5-2B-Base1.9B0.81170.86094.66
    Gemma 3 1B pt999.9M0.83990.87786.81
    Qwen3-0.6B-Base596M0.89920.93212.78

    Runs

    Published · 83 runs

    All Sansar (Sanskrit) runs. Column headers sort the table.
    F11
    f11_slp1_uni8k_d700m_plus_all_v2_4x
    8 Oct 2026 d700m
    704.1M params
    plus_all v2 slice
    489.9M words · 4×
    129,096 71.48 $133.97 Winner 0.5262 0.5507
    F10
    f10_slp1_uni8k_d700m_plus_all_v2_3x
    5 Oct 2026 d700m
    704.1M params
    plus_all v2 slice
    489.9M words · 3×
    96,822 40.45 $52.59 Winner 0.5307 0.5541
    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
    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

    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.