f1_slp1_uni8k_d125m
d125m (91.5M parameters) on rebuild 3 slice (414M words, 1 passes), trained on RTX 3060 12GB (Home GPU). Started 13 Sep 2026, ended 14 Sep 2026.
Hypothesis
Doubling the model to 125M on the same slice moves held-out bits per byte well below F0.
What we learned
Pooled 0.6133: only 2.4% better for twice the model, where the corpus change had bought 10.5%. The Gita still leaked through partial copies.
Scores
Bits per byte, lower is better; change against the parent run, F0.
ex-Gītā (headline)
0.6273
bits per byte
−2.5%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6132
bits per byte
−2.4%Validation split
0.6096
bits per byte
−2.6%| Test set | F1 | F0 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6255 | 0.6344 | −1.4% |
| Bhagavad-gītā (memorisation) | 0.2439 | 0.2284 | +6.8% |
| Out of domain | 0.621 | 0.639 | −2.8% |
| Prose | 0.6107 | 0.6238 | −2.1% |
| Vedic (Ṛgveda) | 1.0917 | 1.1091 | −1.6% |
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
- 91,491,072
- Outside embeddings
- 84,953,856
- Layers · heads · width
- 12 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- rebuild 3 slice
- Words
- 414,000,000
- Training tokens
- 1,335,968,982
- Passes
- 1 passes
- Tokens seen
- 1,335,951,360
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,180 / 27,180
- GPU hours
- 10.78
- Cost
- —
- Spot restarts
- —
Lineage
homed125m