f5_slp1_uni8k_d125m_plus_clean_2x
d125m (97.2M parameters) on plus_clean slice (448.9M words, 1.84 passes), trained on RTX 3060 12GB (Home GPU). Started 1 Oct 2026, ended 2 Oct 2026.
Hypothesis
A second pass over the same clean data still helps a 125M model.
What we learned
Yes: ex-Gita 0.6039, 1.7% better than one pass. Released as sansar-125m v0.1.0.
Scores
Bits per byte, lower is better; change against the parent run, F6-clean.
ex-Gītā (headline)
0.6039
bits per byte
−1.7%ex-Gītā, clean_v1
0.6254
bits per byte
Pooled, all five sets
0.592
bits per byte
−2.0%Validation split
0.5727
bits per byte
−3.0%| Test set | F6-clean-2x | F6-clean (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6052 | 0.6175 | −2.0% |
| Bhagavad-gītā (memorisation) | 0.279 | 0.3251 | −14.2% |
| Out of domain | 0.6052 | 0.6157 | −1.7% |
| Prose | 0.5859 | 0.5962 | −1.7% |
| Vedic (Ṛgveda) | 0.7833 | 0.7919 | −1.1% |
Curves
Training loss
Cross-entropy per token, by step. The first few percent of the run, far higher, run off the top; hover or the table has every value.
Held-out bits per byte
The validation split, evaluated during training, by step. Lower is better.
Throughput
Tokens per second, by step.
Model
- Preset
- d125m
- Parameters
- 97,241,856
- Outside embeddings
- 84,953,856
- Layers · heads · width
- 12 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_clean slice
- Words
- 448,900,000
- Training tokens
- 1,444,720,245
- Passes
- 1.84 passes
- Tokens seen
- 2,664,628,224
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 54,212 / 54,212
- GPU hours
- 24.74
- Cost
- —
- Spot restarts
- —
Lineage
homed125m