f1b_slp1_uni8k_d125m
d125m (91.5M parameters) on rebuild 3 slice (414M words, 1 passes), trained on RTX 3060 12GB (Home GPU). Started 14 Sep 2026, ended 14 Sep 2026.
Hypothesis
Masking every 28-character window of held-out text inside the training data removes the remaining leak.
What we learned
ex-Gita 0.6372, higher than F1 by design: the masking removed soft contamination. The honest baseline; ex-Gita is the headline number from here on.
Scores
Bits per byte, lower is better; change against the parent run, F1.
ex-Gītā (headline)
0.6372
bits per byte
+1.6%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6271
bits per byte
+2.3%Validation split
0.6124
bits per byte
+0.5%| Test set | F1b | F1 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.633 | 0.6255 | +1.2% |
| Bhagavad-gītā (memorisation) | 0.3597 | 0.2439 | +47.5% |
| Out of domain | 0.6351 | 0.621 | +2.3% |
| Prose | 0.6114 | 0.6107 | +0.1% |
| Vedic (Ṛgveda) | 1.0743 | 1.0917 | −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
- 1,332,329,708
- Passes
- 1 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 10.75
- Cost
- —
- Spot restarts
- —
Lineage
homed125m