e14_slp1_uni16k
d60m (69.3M 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 0.7030, within 0.2% of 8k: better on the Gita and prose, clearly worse on Vedic.
Scores
Bits per byte, lower is better; change against the parent run, E9-16k.
ex-Gītā (headline)
0.7134
bits per byte
−5.6%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.703
bits per byte
−5.8%Validation split
0.6366
bits per byte
−6.5%| Test set | E14-16k | E9-16k (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6711 | 0.7139 | −6.0% |
| Bhagavad-gītā (memorisation) | 0.4277 | 0.5075 | −15.7% |
| Out of domain | 0.7121 | 0.7584 | −6.1% |
| Prose | 0.6847 | 0.7246 | −5.5% |
| Vedic (Ṛgveda) | 1.3431 | 1.2575 | +6.8% |
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
- d60m
- Parameters
- 69,317,376
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- 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
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 7,151 / 7,151
- GPU hours
- 2.1
- Cost
- —
- Spot restarts
- —
Lineage
homed60m