f5_slp1_uni8k_d125m_curated
d125m (97.2M parameters) on curated e-text subset (173M words), trained on RTX 3060 12GB (Home GPU). Started 26 Sep 2026, ended 27 Sep 2026.
Hypothesis
A smaller hand-curated e-text subset (173M words) beats the full slice at equal compute.
What we learned
No: ex-Gita 0.6575 vs F4 0.6189 (+6.2%). Fewer words seen more often lose to more varied text.
Scores
Bits per byte, lower is better; change against the parent run, F4.
ex-Gītā (headline)
0.6575
bits per byte
+6.2%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6465
bits per byte
+6.3%Validation split
0.5554
bits per byte
−5.8%| Test set | F5-curated | F4 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6094 | 0.6204 | −1.8% |
| Bhagavad-gītā (memorisation) | 0.3587 | 0.3271 | +9.7% |
| Out of domain | 0.6589 | 0.6158 | +7.0% |
| Prose | 0.6268 | 0.5974 | +4.9% |
| Vedic (Ṛgveda) | 1.2378 | 1.0098 | +22.6% |
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
- curated e-text subset
- Words
- 173,000,000
- Training tokens
- —
- Passes
- —
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 12.34
- Cost
- —
- Spot restarts
- —
Lineage
homed125m