kg_e9s2_slp1_uni8k
d20m (22.7M 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 0.7597 vs 0.7576 for 16k on the same hardware: the gap stays within 0.3%, so the tie holds.
Scores
Bits per byte, lower is better; change against the parent run, E9-8k.
ex-Gītā (headline)
0.7681
bits per byte
+1.7%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7597
bits per byte
+1.8%Validation split
0.6993
bits per byte
+1.8%| Test set | E9-8k s2 | E9-8k (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.7304 | 0.7191 | +1.6% |
| Bhagavad-gītā (memorisation) | 0.5396 | 0.5202 | +3.7% |
| Out of domain | 0.7689 | 0.7551 | +1.8% |
| Prose | 0.7408 | 0.7295 | +1.5% |
| Vedic (Ṛgveda) | 1.2749 | 1.2677 | +0.6% |
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
- d20m
- Parameters
- 22,741,632
- Outside embeddings
- 19,473,024
- Layers · heads · width
- 11 · 6 · 384
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 656,098,446
- Passes
- 0.59 passes
- Tokens seen
- 387,858,432
Compute
- GPU
- T4 16GB
- Where
- Kaggle
- Steps
- 7,891 / 7,891
- GPU hours
- 1.69
- Cost
- Free
- Spot restarts
- —
Lineage
kagglet4d20m