i9small_ctx1024_d60m
d60m (68.9M parameters) on plus_clean subset, trained on RTX 3060 12GB (Home GPU). Started 3 Oct 2026, ended 3 Oct 2026.
Hypothesis
Training at context 1,024 beats 512 at the same number of tokens.
What we learned
Tie: ex-Gita 0.6856 vs 0.6825 at the standard window and 0.6817 at its own. Not worth twice the attention cost at this size; 512 kept.
Scores
Bits per byte, lower is better; change against the parent run, E-CTX base.
ex-Gītā (headline)
0.6856
bits per byte
+0.5%ex-Gītā, clean_v1
—
bits per byte
Pooled, all five sets
0.6778
bits per byte
+0.5%Validation split
0.6666
bits per byte
−0.3%| Test set | E-CTX 1024 | E-CTX base (parent) | Change |
|---|---|---|---|
| DCS gold (classical) | 0.6869 | 0.6795 | +1.1% |
| Bhagavad-gītā (memorisation) | 0.4708 | 0.4622 | +1.9% |
| Out of domain | 0.6889 | 0.686 | +0.4% |
| Prose | 0.6564 | 0.6547 | +0.3% |
| Vedic (Ṛgveda) | 0.9388 | 0.9254 | +1.4% |
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
- 68,924,160
- Outside embeddings
- 56,636,160
- Layers · heads · width
- 8 · 12 · 768
- Tokenizer
- SLP1 unigram 8k
Data
- Slice
- plus_clean subset
- Words
- —
- Training tokens
- 325,604,238
- Passes
- 1 passes
- Tokens seen
- 325,582,848
Compute
- GPU
- RTX 3060 12GB
- Where
- Home GPU
- Steps
- 6,624 / 6,624
- GPU hours
- 2.08
- Cost
- —
- Spot restarts
- —
Lineage
homed60m