f5_slp1_uni8k_d125m_plus_t10
d125m (97.2M parameters) on plus_clean + additions (509.3M words, 0.8 passes), trained on RTX 3060 12GB (Home GPU). Started 2 Oct 2026, ended 3 Oct 2026.
Hypothesis
About 60M words of new commentary-register text help at equal compute.
What we learned
Flat: ex-Gita 0.6135 vs 0.6145 (-0.16%). More words of a different register are not the lever at 125M; model size is.
Scores
Bits per byte, lower is better; change against the parent run, F6-clean.
ex-Gītā (headline)
0.6135
bits per byte
−0.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6036
bits per byte
−0.0%Validation split
0.5978
bits per byte
+1.2%| Test set | F8-small | F6-clean (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6212 | 0.6175 | +0.6% |
| Bhagavad-gītā (memorisation) | 0.3434 | 0.3251 | +5.6% |
| Out of domain | 0.6147 | 0.6157 | −0.2% |
| Prose | 0.5942 | 0.5962 | −0.3% |
| Vedic (Ṛgveda) | 0.7835 | 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 + additions
- Words
- 509,300,000
- Training tokens
- 1,659,569,868
- Passes
- 0.8 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 12.3
- Cost
- —
- Spot restarts
- —
Lineage
homed125m