e14_slp1_uni8k
d60m (63.2M parameters) on E8 screening slice, trained on RTX 3060 12GB (Home GPU). Started 12 Sep 2026, ended 12 Sep 2026.
Hypothesis
The 8k/16k vocabulary tie survives a 3x larger model (60M).
What we learned
Pooled 0.7016, 6% better than its 20M twin and within 0.2% of 16k; 8k is better on Vedic. Supported freezing 8k.
Scores
Bits per byte, lower is better; change against the parent run, E9-8k.
ex-Gītā (headline)
0.7117
bits per byte
−5.8%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.7016
bits per byte
−6.0%Validation split
0.6471
bits per byte
−5.8%| Test set | E14-8k | E9-8k (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6737 | 0.7191 | −6.3% |
| Bhagavad-gītā (memorisation) | 0.4374 | 0.5202 | −15.9% |
| Out of domain | 0.7097 | 0.7551 | −6.0% |
| Prose | 0.6868 | 0.7295 | −5.9% |
| Vedic (Ṛgveda) | 1.292 | 1.2677 | +1.9% |
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
- d60m
- Parameters
- 63,173,376
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- E8 screening slice
- Words
- —
- Training tokens
- 656,098,446
- Passes
- 0.59 passes
- Tokens seen
- 387,907,584
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,892 / 7,892
- GPU hours
- 2.13
- Cost
- —
- Spot restarts
- —
Lineage
homed60m