kg_e9s2_slp1_uni16k
d20m (24M parameters) on E8 screening slice, trained on T4 16GB (Kaggle). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
The 8k/16k vocabulary tie survives a second seed (on a free Kaggle T4).
What we learned
Pooled 0.7576. Across two seeds and two scales the 8k/16k gap never exceeds 0.3%, so 8k was frozen for its smaller table.
Scores
Bits per byte, lower is better; change against the parent run, E9-16k.
ex-Gītā (headline)
0.766
bits per byte
+1.4%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7576
bits per byte
+1.5%Validation split
0.6865
bits per byte
+0.9%| Test set | E9-16k s2 | E9-16k (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7251 | 0.7139 | +1.6% |
| Bhagavad-gītā (memorisation) | 0.5347 | 0.5075 | +5.4% |
| Out of domain | 0.7683 | 0.7584 | +1.3% |
| Prose | 0.7377 | 0.7246 | +1.8% |
| Vedic (Ṛgveda) | 1.2478 | 1.2575 | −0.8% |
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
- d20m
- Parameters
- 24,043,392
- Outside embeddings
- 17,702,784
- Layers · heads · width
- 10 · 6 · 384
- Tokenizer
- SLP1 unigram 16k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 594,539,065
- Passes
- 0.59 passes
- Tokens seen
- 351,485,952
Compute
- GPU
- T4 16GB
- Where
- Kaggle
- Steps
- 7,151 / 7,151
- GPU hours
- 1.62
- Cost
- Free
- Spot restarts
- —
Lineage
kagglet4d20m