f5_slp1_uni8k_d125m_filtered_v2
d125m (97.2M parameters) on line-filtered slice v2, trained on RTX 3060 12GB (Home GPU). Started 27 Sep 2026, ended 27 Sep 2026.
Hypothesis
The same line filter with an accent-blind vocabulary check keeps more of the good lines.
What we learned
No: ex-Gita 0.6256, still behind F4. Line filtering was dropped.
Scores
Bits per byte, lower is better; change against the parent run, F5-filtered.
ex-Gītā (headline)
0.6256
bits per byte
+0.3%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6146
bits per byte
+0.3%Validation split
0.5683
bits per byte
−1.0%| Test set | F5-filtered v2 | F5-filtered (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6172 | 0.6162 | +0.2% |
| Bhagavad-gītā (memorisation) | 0.3238 | 0.3238 | ±0.0% |
| Out of domain | 0.6246 | 0.6217 | +0.5% |
| Prose | 0.5991 | 0.5982 | +0.2% |
| Vedic (Ṛgveda) | 1.0519 | 1.0737 | −2.0% |
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
- line-filtered slice v2
- Words
- —
- Training tokens
- —
- Passes
- —
- Tokens seen
- 1,332,314,112
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 27,106 / 27,106
- GPU hours
- 12.3
- Cost
- —
- Spot restarts
- —
Lineage
homed125m