f5_slp1_uni8k_d125m_plus_clean
d125m (97.2M parameters) on plus_clean slice (448.9M words, 0.92 passes), trained on RTX 3060 12GB (Home GPU). Started 30 Sep 2026, ended 1 Oct 2026.
Hypothesis
The F6 additions, with every held-out overlap filtered out, still help.
What we learned
Yes: ex-Gita 0.6145, 0.7% better than F4, almost all from Vedic (21% better). A split by overlap showed the Vedic gain is genuine register learning.
Scores
Bits per byte, lower is better; change against the parent run, F4.
ex-Gītā (headline)
0.6145
bits per byte
−0.7%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6039
bits per byte
−0.7%Validation split
0.5906
bits per byte
+0.1%| Test set | F6-clean | F4 (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6175 | 0.6204 | −0.5% |
| Bhagavad-gītā (memorisation) | 0.3251 | 0.3271 | −0.6% |
| Out of domain | 0.6157 | 0.6158 | −0.0% |
| Prose | 0.5962 | 0.5974 | −0.2% |
| Vedic (Ṛgveda) | 0.7919 | 1.0098 | −21.6% |
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
- plus_clean slice
- Words
- 448,900,000
- Training tokens
- 1,444,720,245
- Passes
- 0.92 passes
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 12.35
- Cost
- —
- Spot restarts
- —
Lineage
homed125m